* [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
@ 2023-05-04 13:25 juzhe.zhong
2023-05-05 23:41 ` 钟居哲
` (3 more replies)
0 siblings, 4 replies; 25+ messages in thread
From: juzhe.zhong @ 2023-05-04 13:25 UTC (permalink / raw)
To: gcc-patches; +Cc: richard.sandiford, rguenther, Ju-Zhe Zhong
From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
This patch is fixing V3 patch:
https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
Fix issues according to Richard Sandiford && Richard Biener.
1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
2. Support multiple-rgroup for non-SLP auto-vectorization.
For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
_36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
First length (MIN (X, VF/N)):
loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
Second length (X - MIN (X, 1 * VF/N)):
loop_len_16 = _36 - loop_len_15;
Third length (X - MIN (X, 2 * VF/N)):
_38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
loop_len_17 = _36 - _38;
Forth length (X - MIN (X, 3 * VF/N)):
_39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
loop_len_18 = _36 - _39;
The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
I found it's hard to adjust length correctly according to SELECT_VL.
So, this patch we only use SELECT_VL for single-rgroup with single length control.
3. Fix document of select_vl for Richard Biener (remove mode N).
4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
More comments from Richard Biener:
>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
the loop control counter never underflow zero.
>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>> the above scheme should also work for single rgroups, no?
>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>> .WHILE_LEN is a pure optimization?
Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
target adjust any length = INRANGE (0, min (n, vf)) each iteration.
Let me known if I missed something for the V3 patch.
Thanks.
---
gcc/cfgloopmanip.cc | 2 +-
gcc/doc/md.texi | 34 +++
gcc/gimple-loop-interchange.cc | 2 +-
gcc/internal-fn.def | 1 +
gcc/optabs.def | 1 +
gcc/tree-ssa-loop-ivcanon.cc | 2 +-
gcc/tree-ssa-loop-ivopts.cc | 2 +-
gcc/tree-ssa-loop-manip.cc | 18 +-
gcc/tree-ssa-loop-manip.h | 4 +-
gcc/tree-vect-data-refs.cc | 8 +-
gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
gcc/tree-vect-loop.cc | 32 ++-
gcc/tree-vect-stmts.cc | 89 +++++++-
gcc/tree-vectorizer.h | 4 +-
14 files changed, 535 insertions(+), 38 deletions(-)
diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
index 0e3ad8ed742..6e09dcbb0b1 100644
--- a/gcc/cfgloopmanip.cc
+++ b/gcc/cfgloopmanip.cc
@@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
}
gsi = gsi_last_bb (loop_header);
- create_iv (initial_value, stride, iv, loop, &gsi, false,
+ create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
iv_before, iv_after);
/* Insert loop exit condition. */
diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
index cc4a93a8763..99cf0cdbdca 100644
--- a/gcc/doc/md.texi
+++ b/gcc/doc/md.texi
@@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
@end smallexample
+@cindex @code{select_vl@var{m}} instruction pattern
+@item @code{select_vl@var{m}}
+Set operand 0 to the number of active elements in vector will be updated value.
+operand 1 is the total elements need to be updated value.
+operand 2 is the vectorization factor.
+The value of operand 0 is target dependent and flexible in each iteration.
+The operation of this pattern can be:
+
+@smallexample
+Case 1:
+operand0 = MIN (operand1, operand2);
+operand2 can be const_poly_int or poly_int related to vector mode size.
+Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
+that we can reduce a use of general purpose register.
+
+In this case, only the last iteration of the loop is partial iteration.
+@end smallexample
+
+@smallexample
+Case 2:
+if (operand1 <= operand2)
+ operand0 = operand1;
+else if (operand1 < 2 * operand2)
+ operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
+else
+ operand0 = operand2;
+
+This case will evenly distribute work over the last 2 iterations of a stripmine loop.
+@end smallexample
+
+The output of this pattern is not only used as IV of loop control counter, but also
+is used as the IV of address calculation with multiply/shift operation. This allow
+us dynamic adjust the number of elements is processed in each iteration of the loop.
+
@cindex @code{check_raw_ptrs@var{m}} instruction pattern
@item @samp{check_raw_ptrs@var{m}}
Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
index 1b77bfd46b2..e5590374e59 100644
--- a/gcc/gimple-loop-interchange.cc
+++ b/gcc/gimple-loop-interchange.cc
@@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
tree var_before, var_after;
tree base = unshare_expr (iv->init_expr);
tree step = unshare_expr (iv->step);
- create_iv (base, step, SSA_NAME_VAR (iv->var),
+ create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
tgt.m_loop, &incr_pos, false, &var_before, &var_after);
bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
index 7fe742c2ae7..6f6fa7d37f9 100644
--- a/gcc/internal-fn.def
+++ b/gcc/internal-fn.def
@@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
+DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
check_raw_ptrs, check_ptrs)
DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
diff --git a/gcc/optabs.def b/gcc/optabs.def
index 695f5911b30..b637471b76e 100644
--- a/gcc/optabs.def
+++ b/gcc/optabs.def
@@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
OPTAB_D (len_load_optab, "len_load_$a")
OPTAB_D (len_store_optab, "len_store_$a")
+OPTAB_D (select_vl_optab, "select_vl$a")
diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
index e41ec73a52a..d4e113c44df 100644
--- a/gcc/tree-ssa-loop-ivcanon.cc
+++ b/gcc/tree-ssa-loop-ivcanon.cc
@@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
niter,
build_int_cst (type, 1));
incr_at = gsi_last_bb (in->src);
- create_iv (niter,
+ create_iv (niter, PLUS_EXPR,
build_int_cst (type, -1),
NULL_TREE, loop,
&incr_at, false, var_before, &var);
diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
index 78e8cbc75b5..8640fe2d487 100644
--- a/gcc/tree-ssa-loop-ivopts.cc
+++ b/gcc/tree-ssa-loop-ivopts.cc
@@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
base = unshare_expr (cand->iv->base);
- create_iv (base, unshare_expr (cand->iv->step),
+ create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
cand->var_before, data->current_loop,
&incr_pos, after, &cand->var_before, &cand->var_after);
}
diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
index 909b705d00d..5abca64379e 100644
--- a/gcc/tree-ssa-loop-manip.cc
+++ b/gcc/tree-ssa-loop-manip.cc
@@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
so that we can free them all at once. */
static bitmap_obstack loop_renamer_obstack;
-/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
+/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
+ If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
+ If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
It is expected that neither BASE nor STEP are shared with other expressions
(unless the sharing rules allow this). Use VAR as a base var_decl for it
(if NULL, a new temporary will be created). The increment will occur at
@@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
VAR_AFTER (unless they are NULL). */
void
-create_iv (tree base, tree step, tree var, class loop *loop,
- gimple_stmt_iterator *incr_pos, bool after,
+create_iv (tree base, tree_code code, tree step, tree var,
+ class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
tree *var_before, tree *var_after)
{
gassign *stmt;
@@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
tree initial, step1;
gimple_seq stmts;
tree vb, va;
- enum tree_code incr_op = PLUS_EXPR;
+ /* The code can only be PLUS_EXPR or MINUS_EXPR. */
+ gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
+ tree_code incr_op = code;
edge pe = loop_preheader_edge (loop);
if (var != NULL_TREE)
@@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
tree ctr_before, ctr_after;
gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
exit_if = as_a <gcond *> (gsi_stmt (bsi));
- create_iv (exit_base, exit_step, NULL_TREE, loop,
+ create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
&bsi, false, &ctr_before, &ctr_after);
gimple_cond_set_code (exit_if, exit_cmp);
gimple_cond_set_lhs (exit_if, ctr_after);
@@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
gsi = gsi_last_bb (loop->latch);
else
gsi = gsi_last_nondebug_bb (loop->header);
- create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
- loop, &gsi, bump_in_latch, &var_before, NULL);
+ create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
+ NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
rewrite_all_phi_nodes_with_iv (loop, var_before);
diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
index d49273a3987..bda09f51d56 100644
--- a/gcc/tree-ssa-loop-manip.h
+++ b/gcc/tree-ssa-loop-manip.h
@@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
typedef void (*transform_callback)(class loop *, void *);
-extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
- bool, tree *, tree *);
+extern void create_iv (tree, tree_code, tree, tree, class loop *,
+ gimple_stmt_iterator *, bool, tree *, tree *);
extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
extern void verify_loop_closed_ssa (bool, class loop * = NULL);
diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index 6721ab6efc4..5c9103b16e5 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (aggr_ptr_init,
+ create_iv (aggr_ptr_init, PLUS_EXPR,
fold_convert (aggr_ptr_type, iv_step),
aggr_ptr, loop, &incr_gsi, insert_after,
&indx_before_incr, &indx_after_incr);
@@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
{
standard_iv_increment_position (containing_loop, &incr_gsi,
&insert_after);
- create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
- containing_loop, &incr_gsi, insert_after, &indx_before_incr,
- &indx_after_incr);
+ create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
+ aggr_ptr, containing_loop, &incr_gsi, insert_after,
+ &indx_before_incr, &indx_after_incr);
incr = gsi_stmt (incr_gsi);
/* Copy the points-to information if it exists. */
diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
index 44bd5f2c805..d63ded5d4f0 100644
--- a/gcc/tree-vect-loop-manip.cc
+++ b/gcc/tree-vect-loop-manip.cc
@@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
return false;
}
+/* Try to use permutes to define the lens in DEST_RGM using the lens
+ in SRC_RGM, given that the former has twice as many lens as the
+ latter. Return true on success, adding any new statements to SEQ. */
+
+static bool
+vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
+ rgroup_controls *dest_rgm,
+ rgroup_controls *src_rgm)
+{
+ tree ctrl_type = dest_rgm->type;
+ poly_uint64 nitems_per_ctrl
+ = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
+ if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
+ {
+ for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
+ {
+ tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
+ tree dest = dest_rgm->controls[i];
+ gassign *stmt;
+ if (i == 0)
+ {
+ /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
+ tree factor = build_int_cst (iv_type, nitems_per_ctrl);
+ stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
+ gimple_seq_add_stmt (seq, stmt);
+ }
+ else
+ {
+ /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
+ tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
+ tree temp = make_ssa_name (iv_type);
+ stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
+ gimple_seq_add_stmt (seq, stmt);
+ stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
+ gimple_seq_add_stmt (seq, stmt);
+ }
+ }
+ return true;
+ }
+ return false;
+}
+
/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
for all the rgroup controls in RGC and return a control that is nonzero
when the loop needs to iterate. Add any new preheader statements to
@@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
gimple_stmt_iterator incr_gsi;
bool insert_after;
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
- &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
+ create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
+ loop, &incr_gsi, insert_after, &index_before_incr,
+ &index_after_incr);
tree zero_index = build_int_cst (compare_type, 0);
tree test_index, test_limit, first_limit;
@@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
return next_ctrl;
}
+/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
+ for all the rgroup controls in RGC and return a control that is nonzero
+ when the loop needs to iterate. Add any new preheader statements to
+ PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
+
+ RGC belongs to loop LOOP. The loop originally iterated NITERS
+ times and has been vectorized according to LOOP_VINFO.
+
+ Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
+ to TEST_LIMIT - bias.
+
+ In vect_set_loop_controls_by_select_vl, we are iterating from start at
+ IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
+ IFN_SELECT_VL pattern.
+
+ 1. Single rgroup, the Gimple IR should be:
+
+ # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
+ # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
+ # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
+ # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
+ # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
+ _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
+ ivtmp_15 = _28 * 4;
+ vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
+ _1 = B[i_10];
+ .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
+ i_7 = i_10 + 1;
+ vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
+ vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
+ ivtmp_27 = ivtmp_26 - _28;
+ if (ivtmp_27 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 7>; [16.67%]
+
+ Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
+ data reference pointer IV.
+
+ 1). The result of .SELECT_VL:
+ _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
+ The _28 is not necessary to be VF in any iteration, instead, we allow
+ _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
+ pattern allows target have various flexible optimizations in vector
+ loop iterations. Target like RISC-V has special application vector
+ length calculation instruction which will distribute even workload
+ in the last 2 iterations.
+
+ Other example is that we can allow even generate _28 <= VF / 2 so
+ that some machine can run vector codes in low power mode.
+
+ 2). Loop control IV:
+ ivtmp_27 = ivtmp_26 - _28;
+ if (ivtmp_27 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 7>; [16.67%]
+
+ This is the saturating-subtraction towards zero, the outcome of
+ .SELECT_VL wil make ivtmp_27 never underflow zero.
+
+ 3). Data reference pointer IV:
+ ivtmp_15 = _28 * 4;
+ vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
+ vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
+
+ The pointer IV is adjusted accurately according to the .SELECT_VL.
+
+ 2. Multiple rgroup, the Gimple IR should be:
+
+ # i_23 = PHI <i_20(6), 0(11)>
+ # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
+ # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
+ # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
+ # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ _1 = i_23 * 2;
+ _2 = (long unsigned int) _1;
+ _3 = _2 * 2;
+ _4 = f_15(D) + _3;
+ _5 = _2 + 1;
+ _6 = _5 * 2;
+ _7 = f_15(D) + _6;
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ _8 = (long unsigned int) i_23;
+ _9 = _8 * 4;
+ _10 = d_18(D) + _9;
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+ i_20 = i_23 + 1;
+ vectp_f.8_52 = vectp_f.8_56 + 16;
+ vectp_d.10_60 = vectp_d.10_63 + 16;
+ ivtmp_74 = ivtmp_73 - _75;
+ ivtmp_71 = ivtmp_70 - _72;
+ if (ivtmp_74 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 13>; [16.67%]
+
+ Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
+ rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
+ iteration amount for mutiple rgroups.
+
+ The analysis of the flow of multiple rgroups:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ ...
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+
+ We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
+ to be processed in each iteration.
+
+ The related STOREs:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+ Since these 2 STOREs store 2 vectors that the second vector is half elements
+ of the first vector. So the length of second STORE will be _64 = _72 / 2;
+ It's similar to the VIEW_CONVERT of handling masks in SLP.
+
+ 3. Multiple rgroups for non-SLP auto-vectorization.
+
+ # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
+ # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
+ # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
+ _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+ _29 = (void *) ivtmp.35_10;
+ _7 = &MEM <vector([4,4]) int> [(int *)_29];
+ vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
+ _33 = _29 + POLY_INT_CST [16, 16];
+ _34 = &MEM <vector([4,4]) int> [(int *)_33];
+ vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
+ vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
+ _30 = (void *) ivtmp.36_2;
+ _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
+ .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
+ ivtmp_27 = ivtmp_26 - _28;
+ ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
+ ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
+ if (ivtmp_27 != 0)
+ goto <bb 4>; [83.33%]
+ else
+ goto <bb 5>; [16.67%]
+
+ The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+
+ The length of first half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+
+ The length of second half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+
+ 1). _28 always <= POLY_INT_CST [8, 8].
+ 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
+ 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
+*/
+
+static tree
+vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
+ gimple_seq *preheader_seq,
+ gimple_seq *header_seq,
+ rgroup_controls *rgc, tree niters,
+ unsigned int controls_length)
+{
+ tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
+ /* We are not allowing masked approach in SELECT_VL. */
+ gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
+
+ tree ctrl_type = rgc->type;
+ unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
+ poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
+ poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+
+ /* Calculate the maximum number of item values that the rgroup
+ handles in total, the number that it handles for each iteration
+ of the vector loop. */
+ tree nitems_total = niters;
+ if (nitems_per_iter != 1)
+ {
+ /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
+ these multiplications don't overflow. */
+ tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
+ nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
+ nitems_total, compare_factor);
+ }
+
+ /* Convert the comparison value to the IV type (either a no-op or
+ a promotion). */
+ nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
+
+ /* Create an induction variable that counts the number of items
+ processed. */
+ tree index_before_incr, index_after_incr;
+ gimple_stmt_iterator incr_gsi;
+ bool insert_after;
+ standard_iv_increment_position (loop, &incr_gsi, &insert_after);
+
+ /* Test the decremented IV, which will never underflow 0 since we have
+ IFN_SELECT_VL to gurantee that. */
+ tree test_limit = nitems_total;
+
+ /* Provide a definition of each control in the group. */
+ tree ctrl;
+ unsigned int i;
+ FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
+ {
+ /* Previous controls will cover BIAS items. This control covers the
+ next batch. */
+ poly_uint64 bias = nitems_per_ctrl * i;
+ tree bias_tree = build_int_cst (iv_type, bias);
+
+ /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
+ BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
+ control and adjust the bound down by BIAS. */
+ tree this_test_limit = test_limit;
+ if (i != 0)
+ {
+ this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
+ this_test_limit, bias_tree);
+ this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
+ this_test_limit, bias_tree);
+ }
+
+ /* Create decrement IV. */
+ create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
+ insert_after, &index_before_incr, &index_after_incr);
+
+ poly_uint64 final_vf = vf * nitems_per_iter;
+ tree vf_step = build_int_cst (iv_type, final_vf);
+ tree res_len;
+ if (nitems_per_iter != 1 || controls_length != 1)
+ {
+ /* For SLP, we can't allow non-VF number of elements to be processed
+ in non-final iteration. We force the number of elements to be
+ processed in each non-final iteration is VF elements. If we allow
+ non-VF elements processing in non-final iteration will make SLP too
+ complicated and produce inferior codegen.
+
+ For example:
+
+ If non-final iteration process VF elements.
+
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
+ ...
+
+ If non-final iteration process non-VF elements.
+
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
+ if (_71 % 2 == 0)
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
+ else
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
+ ...
+
+ This is the simple case of 2-elements interleaved vector SLP. We
+ consider other interleave vector, the situation will become more
+ complicated. */
+ res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
+ index_before_incr, vf_step);
+ }
+ else
+ {
+ res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
+ index_before_incr, vf_step);
+ }
+ gassign *assign = gimple_build_assign (ctrl, res_len);
+ gimple_seq_add_stmt (header_seq, assign);
+ }
+
+ return index_after_incr;
+}
+
/* Set up the iteration condition and rgroup controls for LOOP, given
that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
@@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
+ bool use_vl_p = !use_masks_p
+ && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
+ OPTIMIZE_FOR_SPEED);
unsigned int compare_precision = TYPE_PRECISION (compare_type);
tree orig_niters = niters;
@@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
continue;
}
+ if (use_vl_p && rgc->controls.length () != 1)
+ {
+ rgroup_controls *sub_rgc
+ = &(*controls)[nmasks / rgc->controls.length () - 1];
+ if (!sub_rgc->controls.is_empty ()
+ && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
+ sub_rgc))
+ continue;
+ }
+
/* See whether zero-based IV would ever generate all-false masks
or zero length before wrapping around. */
bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
/* Set up all controls for this group. */
- test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
- &preheader_seq,
- &header_seq,
- loop_cond_gsi, rgc,
- niters, niters_skip,
- might_wrap_p);
+ if (use_vl_p)
+ test_ctrl
+ = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
+ &preheader_seq, &header_seq,
+ rgc, niters, controls->length ());
+ else
+ test_ctrl
+ = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
+ &header_seq, loop_cond_gsi, rgc,
+ niters, niters_skip,
+ might_wrap_p);
}
/* Emit all accumulated statements. */
@@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
}
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (init, step, NULL_TREE, loop,
+ create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
&incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
true, NULL_TREE, true,
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index 6ea0f21fd13..d7cf92576d1 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
gimple_stmt_iterator incr_gsi;
bool insert_after;
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
+ create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
insert_after, &indx_before_incr, &indx_after_incr);
/* Next create a new phi node vector (NEW_PHI_TREE) which starts
@@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
}
/* Given a complete set of length LENS, extract length number INDEX for an
- rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
+ rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
+ Insert any set-up statements before GSI. */
tree
-vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
- unsigned int nvectors, unsigned int index)
+vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
+ vec_loop_lens *lens, unsigned int nvectors, tree vectype,
+ unsigned int index)
{
rgroup_controls *rgl = &(*lens)[nvectors - 1];
bool use_bias_adjusted_len =
LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
/* Populate the rgroup's len array, if this is the first time we've
used it. */
@@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
if (use_bias_adjusted_len)
return rgl->bias_adjusted_ctrl;
+ else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
+ OPTIMIZE_FOR_SPEED))
+ {
+ tree loop_len = rgl->controls[index];
+ poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
+ poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
+ if (maybe_ne (nunits1, nunits2))
+ {
+ /* A loop len for data type X can be reused for data type Y
+ if X has N times more elements than Y and if Y's elements
+ are N times bigger than X's. */
+ gcc_assert (multiple_p (nunits1, nunits2));
+ unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
+ gimple_seq seq = NULL;
+ loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
+ build_int_cst (iv_type, factor));
+ if (seq)
+ gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
+ }
+ return loop_len;
+ }
else
return rgl->controls[index];
}
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 3ad6a7d28d7..26813ab56e9 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
return iv_step;
}
+/* Prepare the pointer IVs which needs to be updated by a variable amount.
+ Such variable amount is the outcome of .SELECT_VL. In this case, we can
+ allow each iteration process the flexible number of elements as long as
+ the number <= vf elments.
+
+ Return data reference according to SELECT_VL.
+ If new statements are needed, insert them before GSI. */
+
+static tree
+get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
+ tree aggr_type, class loop *at_loop, tree offset,
+ tree *dummy, gimple_stmt_iterator *gsi,
+ bool simd_lane_access_p, vec_loop_lens *loop_lens,
+ dr_vec_info *dr_info,
+ vect_memory_access_type memory_access_type)
+{
+ loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
+ tree step = vect_dr_behavior (vinfo, dr_info)->step;
+
+ /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
+ IVs are updated by variable amount but we will support them in the future.
+ */
+ gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
+ && memory_access_type != VMAT_LOAD_STORE_LANES);
+
+ /* When we support SELECT_VL pattern, we dynamic adjust
+ the memory address by .SELECT_VL result.
+
+ The result of .SELECT_VL is the number of elements to
+ be processed of each iteration. So the memory address
+ adjustment operation should be:
+
+ bytesize = GET_MODE_SIZE (element_mode (aggr_type));
+ addr = addr + .SELECT_VL (ARG..) * bytesize;
+ */
+ gimple *ptr_incr;
+ tree loop_len
+ = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
+ tree len_type = TREE_TYPE (loop_len);
+ poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
+ /* Since the outcome of .SELECT_VL is element size, we should adjust
+ it into bytesize so that it can be used in address pointer variable
+ amount IVs adjustment. */
+ tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
+ build_int_cst (len_type, bytesize));
+ if (tree_int_cst_sgn (step) == -1)
+ tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
+ tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
+ gassign *assign = gimple_build_assign (bump, tmp);
+ gsi_insert_before (gsi, assign, GSI_SAME_STMT);
+ return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
+ dummy, gsi, &ptr_incr, simd_lane_access_p,
+ bump);
+}
+
/* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
static bool
@@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
- create_iv (stride_base, ivstep, NULL,
+ create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
loop, &incr_gsi, insert_after,
&offvar, NULL);
incr = gsi_stmt (incr_gsi);
@@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
slp_node, &gs_info, &dataref_ptr,
&vec_offsets);
+ else if (loop_lens && loop_lens->length () == 1
+ && direct_internal_fn_supported_p (
+ IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
+ OPTIMIZE_FOR_SPEED)
+ && memory_access_type != VMAT_INVARIANT)
+ dataref_ptr
+ = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
+ simd_lane_access_p ? loop : NULL,
+ offset, &dummy, gsi,
+ simd_lane_access_p, loop_lens,
+ dr_info, memory_access_type);
else
dataref_ptr
= vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
@@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
else if (loop_lens)
{
tree final_len
- = vect_get_loop_len (loop_vinfo, loop_lens,
- vec_num * ncopies, vec_num * j + i);
+ = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ vec_num * ncopies, vectype,
+ vec_num * j + i);
tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
machine_mode vmode = TYPE_MODE (vectype);
opt_machine_mode new_ovmode
@@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
- create_iv (stride_base, ivstep, NULL,
+ create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
loop, &incr_gsi, insert_after,
&offvar, NULL);
@@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
slp_node, &gs_info, &dataref_ptr,
&vec_offsets);
}
+ else if (loop_lens && loop_lens->length () == 1
+ && direct_internal_fn_supported_p (
+ IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
+ OPTIMIZE_FOR_SPEED)
+ && memory_access_type != VMAT_INVARIANT)
+ dataref_ptr
+ = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
+ at_loop, offset, &dummy, gsi,
+ simd_lane_access_p, loop_lens,
+ dr_info, memory_access_type);
else
dataref_ptr
= vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
@@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
else if (loop_lens && memory_access_type != VMAT_INVARIANT)
{
tree final_len
- = vect_get_loop_len (loop_vinfo, loop_lens,
- vec_num * ncopies,
+ = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ vec_num * ncopies, vectype,
vec_num * j + i);
tree ptr = build_int_cst (ref_type,
align * BITS_PER_UNIT);
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 9cf2fb23fe3..357bc5c7315 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
unsigned int, tree, unsigned int);
extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
tree, unsigned int);
-extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
- unsigned int);
+extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
+ unsigned int, tree, unsigned int);
extern gimple_seq vect_gen_len (tree, tree, tree, tree);
extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
--
2.36.1
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-04 13:25 [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support juzhe.zhong
@ 2023-05-05 23:41 ` 钟居哲
2023-05-09 12:59 ` Richard Sandiford
2023-05-07 15:19 ` Jeff Law
` (2 subsequent siblings)
3 siblings, 1 reply; 25+ messages in thread
From: 钟居哲 @ 2023-05-05 23:41 UTC (permalink / raw)
To: gcc-patches; +Cc: richard.sandiford, rguenther
[-- Attachment #1: Type: text/plain, Size: 45269 bytes --]
Hi, Richards. I would like to give more information about this patch so that it will make this patch easier for you to review.
Currently, I saw we have 3 situations that we need to handle in case of loop control IV in auto-vectorization:
1. Single rgroup loop control (ncopies == 1 && vec_num == 1 so loop_len.length () == 1 or rgc->lengh () == 1)
2. Multiple rgroup for SLP.
3. Multiple rgroup for non-SLP which is Richard Sandiford point out previously (For example, VEC_PACK_TRUNC).
To talk about this patch, let me talk about RVV LLVM implementation first which inspire me to send this patch:
https://reviews.llvm.org/D99750
According to LLVM implementation, they are adding a middle-end IR called "get_vector_length" which has totally
same functionality as "select_vl" in this patch (I call it "while_len" previously, now I rename it as "select_vl" following Richard suggestion).
The LLVM implementation is only let "get_vector_length" calculate the number of elements in single rgroup loop.
For multi rgroup, let's take a look at it:
https://godbolt.org/z/3GP78efTY
void
foo1 (short *__restrict f, int *__restrict d, int n)
{
for (int i = 0; i < n; ++i)
{
f[i * 2 + 0] = 1;
f[i * 2 + 1] = 2;
d[i] = 3;
}
}
RISC-V Clang:
foo1: # @foo1
# %bb.0:
blez a2, .LBB0_8
# %bb.1:
li a3, 16
bgeu a2, a3, .LBB0_3
# %bb.2:
li a6, 0
j .LBB0_6
.LBB0_3:
andi a6, a2, -16
lui a3, 32
addiw a3, a3, 1
vsetivli zero, 8, e32, m2, ta, ma
vmv.v.x v8, a3
vmv.v.i v10, 3
mv a4, a6
mv a5, a1
mv a3, a0
.LBB0_4: # =>This Inner Loop Header: Depth=1
addi a7, a5, 32
addi t0, a3, 32
vsetivli zero, 16, e16, m2, ta, ma
vse16.v v8, (a3)
vse16.v v8, (t0)
vsetivli zero, 8, e32, m2, ta, ma
vse32.v v10, (a5)
vse32.v v10, (a7)
addi a3, a3, 64
addi a4, a4, -16
addi a5, a5, 64
bnez a4, .LBB0_4
# %bb.5:
beq a6, a2, .LBB0_8
.LBB0_6:
slli a3, a6, 2
add a0, a0, a3
addi a0, a0, 2
add a1, a1, a3
sub a2, a2, a6
li a3, 1
li a4, 2
li a5, 3
.LBB0_7: # =>This Inner Loop Header: Depth=1
sh a3, -2(a0)
sh a4, 0(a0)
sw a5, 0(a1)
addi a0, a0, 4
addi a2, a2, -1
addi a1, a1, 4
bnez a2, .LBB0_7
.LBB0_8:
ret
ARM GCC:
foo1:
cmp w2, 0
ble .L1
addvl x4, x0, #1
mov x3, 0
cntb x7
cntb x6, all, mul #2
sbfiz x2, x2, 1, 32
ptrue p0.b, all
mov x5, x2
adrp x8, .LC0
uqdech x5
add x8, x8, :lo12:.LC0
whilelo p1.h, xzr, x5
ld1rw z1.s, p0/z, [x8]
mov z0.s, #3
whilelo p0.h, xzr, x2
.L3:
st1h z1.h, p0, [x0, x3, lsl 1]
st1h z1.h, p1, [x4, x3, lsl 1]
st1w z0.s, p1, [x1, #1, mul vl]
add x3, x3, x7
whilelo p1.h, x3, x5
st1w z0.s, p0, [x1]
add x1, x1, x6
whilelo p0.h, x3, x2
b.any .L3
.L1:
ret
It's very obvious that ARM GCC has much better codegen since RVV LLVM just use SIMD style to handle multi-rgroup SLP auto-vectorization.
Well, I am totally aggree that we should add length stuff in auto-vectorization not only for single rgroup but also multiple rgroup.
However, when I am trying to implement multiple rgroup length for both SLP and non-SLP and testing, turns out it's hard to use select_vl
since "select_vl" pattern allows non-VF flexible length (length <= min (remain,VF)) in any iteration, it's consuming much more operations for
adjust loop controls IV and data reference address point IV than just using "MIN_EXPR".
So for Case 2 && Case 3, I just use MIN_EXPR directly instead of SELECT_VL after my serveral internal testing.
Now base on these situations, we only have "select_vl" for single-rgroup, but multiple-rgroup (both SLP and non-SLP), we just
use MIN_EXPR.
Is it more appropriate that we should remove "select_vl" and just use MIN_EXPR force VF elements in each non-final iteration in single rgroup?
Like the codegen according to RVV ISA example (show as RVV LLVM):
https://repo.hca.bsc.es/epic/z/oynhzP
ASM:
vec_add: # @vec_add
blez a3, .LBB0_3
li a4, 0
.LBB0_2: # %vector.body
sub a5, a3, a4
vsetvli a6, a5, e64, m1, ta, mu ==> change it into a6 = min (a5, VF) && vsetvli zero, a6, e64, m1, ta, mu
slli a7, a4, 3
add a5, a1, a7
vle64.v v8, (a5)
add a5, a2, a7
vle64.v v9, (a5)
vfadd.vv v8, v8, v9
add a7, a7, a0
add a4, a4, a6
vse64.v v8, (a7)
bne a4, a3, .LBB0_2
.LBB0_3: # %for.cond.cleanup
ret
So if we remove "select_vl" just use MIN_EXPR to force VF elements in non-final iteration, we will end up with having like this:
vec_add: # @vec_add
blez a3, .LBB0_3
csrr VF in bytes, vlenb
.LBB0_2: # %vector.body
sub a5, a3, VF in elements
a6 = min (a5, VF in elements)
vsetvli zero, a6, e64, m1, ta, mu
add a5, a1, VF in bytes
vle64.v v8, (a5)
add a5, a2, VF in bytes
vle64.v v9, (a5)
vfadd.vv v8, v8, v9
add a7, a7, a0
vse64.v v8, (a7)
bne a4, a3, .LBB0_2
.LBB0_3: # %for.cond.cleanup
ret
Both codegen are working good for RVV.
Just only the second one can not have the RVV special optimization (even distributed workload in last 2 iterations).
Expecting any suggestions from you.
Thanks.
juzhe.zhong@rivai.ai
From: juzhe.zhong
Date: 2023-05-04 21:25
To: gcc-patches
CC: richard.sandiford; rguenther; Ju-Zhe Zhong
Subject: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
This patch is fixing V3 patch:
https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
Fix issues according to Richard Sandiford && Richard Biener.
1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
2. Support multiple-rgroup for non-SLP auto-vectorization.
For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
_36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
First length (MIN (X, VF/N)):
loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
Second length (X - MIN (X, 1 * VF/N)):
loop_len_16 = _36 - loop_len_15;
Third length (X - MIN (X, 2 * VF/N)):
_38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
loop_len_17 = _36 - _38;
Forth length (X - MIN (X, 3 * VF/N)):
_39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
loop_len_18 = _36 - _39;
The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
I found it's hard to adjust length correctly according to SELECT_VL.
So, this patch we only use SELECT_VL for single-rgroup with single length control.
3. Fix document of select_vl for Richard Biener (remove mode N).
4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
More comments from Richard Biener:
>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
the loop control counter never underflow zero.
>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>> the above scheme should also work for single rgroups, no?
>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>> .WHILE_LEN is a pure optimization?
Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
target adjust any length = INRANGE (0, min (n, vf)) each iteration.
Let me known if I missed something for the V3 patch.
Thanks.
---
gcc/cfgloopmanip.cc | 2 +-
gcc/doc/md.texi | 34 +++
gcc/gimple-loop-interchange.cc | 2 +-
gcc/internal-fn.def | 1 +
gcc/optabs.def | 1 +
gcc/tree-ssa-loop-ivcanon.cc | 2 +-
gcc/tree-ssa-loop-ivopts.cc | 2 +-
gcc/tree-ssa-loop-manip.cc | 18 +-
gcc/tree-ssa-loop-manip.h | 4 +-
gcc/tree-vect-data-refs.cc | 8 +-
gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
gcc/tree-vect-loop.cc | 32 ++-
gcc/tree-vect-stmts.cc | 89 +++++++-
gcc/tree-vectorizer.h | 4 +-
14 files changed, 535 insertions(+), 38 deletions(-)
diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
index 0e3ad8ed742..6e09dcbb0b1 100644
--- a/gcc/cfgloopmanip.cc
+++ b/gcc/cfgloopmanip.cc
@@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
}
gsi = gsi_last_bb (loop_header);
- create_iv (initial_value, stride, iv, loop, &gsi, false,
+ create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
iv_before, iv_after);
/* Insert loop exit condition. */
diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
index cc4a93a8763..99cf0cdbdca 100644
--- a/gcc/doc/md.texi
+++ b/gcc/doc/md.texi
@@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
@end smallexample
+@cindex @code{select_vl@var{m}} instruction pattern
+@item @code{select_vl@var{m}}
+Set operand 0 to the number of active elements in vector will be updated value.
+operand 1 is the total elements need to be updated value.
+operand 2 is the vectorization factor.
+The value of operand 0 is target dependent and flexible in each iteration.
+The operation of this pattern can be:
+
+@smallexample
+Case 1:
+operand0 = MIN (operand1, operand2);
+operand2 can be const_poly_int or poly_int related to vector mode size.
+Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
+that we can reduce a use of general purpose register.
+
+In this case, only the last iteration of the loop is partial iteration.
+@end smallexample
+
+@smallexample
+Case 2:
+if (operand1 <= operand2)
+ operand0 = operand1;
+else if (operand1 < 2 * operand2)
+ operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
+else
+ operand0 = operand2;
+
+This case will evenly distribute work over the last 2 iterations of a stripmine loop.
+@end smallexample
+
+The output of this pattern is not only used as IV of loop control counter, but also
+is used as the IV of address calculation with multiply/shift operation. This allow
+us dynamic adjust the number of elements is processed in each iteration of the loop.
+
@cindex @code{check_raw_ptrs@var{m}} instruction pattern
@item @samp{check_raw_ptrs@var{m}}
Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
index 1b77bfd46b2..e5590374e59 100644
--- a/gcc/gimple-loop-interchange.cc
+++ b/gcc/gimple-loop-interchange.cc
@@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
tree var_before, var_after;
tree base = unshare_expr (iv->init_expr);
tree step = unshare_expr (iv->step);
- create_iv (base, step, SSA_NAME_VAR (iv->var),
+ create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
tgt.m_loop, &incr_pos, false, &var_before, &var_after);
bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
index 7fe742c2ae7..6f6fa7d37f9 100644
--- a/gcc/internal-fn.def
+++ b/gcc/internal-fn.def
@@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
+DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
check_raw_ptrs, check_ptrs)
DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
diff --git a/gcc/optabs.def b/gcc/optabs.def
index 695f5911b30..b637471b76e 100644
--- a/gcc/optabs.def
+++ b/gcc/optabs.def
@@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
OPTAB_D (len_load_optab, "len_load_$a")
OPTAB_D (len_store_optab, "len_store_$a")
+OPTAB_D (select_vl_optab, "select_vl$a")
diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
index e41ec73a52a..d4e113c44df 100644
--- a/gcc/tree-ssa-loop-ivcanon.cc
+++ b/gcc/tree-ssa-loop-ivcanon.cc
@@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
niter,
build_int_cst (type, 1));
incr_at = gsi_last_bb (in->src);
- create_iv (niter,
+ create_iv (niter, PLUS_EXPR,
build_int_cst (type, -1),
NULL_TREE, loop,
&incr_at, false, var_before, &var);
diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
index 78e8cbc75b5..8640fe2d487 100644
--- a/gcc/tree-ssa-loop-ivopts.cc
+++ b/gcc/tree-ssa-loop-ivopts.cc
@@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
base = unshare_expr (cand->iv->base);
- create_iv (base, unshare_expr (cand->iv->step),
+ create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
cand->var_before, data->current_loop,
&incr_pos, after, &cand->var_before, &cand->var_after);
}
diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
index 909b705d00d..5abca64379e 100644
--- a/gcc/tree-ssa-loop-manip.cc
+++ b/gcc/tree-ssa-loop-manip.cc
@@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
so that we can free them all at once. */
static bitmap_obstack loop_renamer_obstack;
-/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
+/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
+ If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
+ If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
It is expected that neither BASE nor STEP are shared with other expressions
(unless the sharing rules allow this). Use VAR as a base var_decl for it
(if NULL, a new temporary will be created). The increment will occur at
@@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
VAR_AFTER (unless they are NULL). */
void
-create_iv (tree base, tree step, tree var, class loop *loop,
- gimple_stmt_iterator *incr_pos, bool after,
+create_iv (tree base, tree_code code, tree step, tree var,
+ class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
tree *var_before, tree *var_after)
{
gassign *stmt;
@@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
tree initial, step1;
gimple_seq stmts;
tree vb, va;
- enum tree_code incr_op = PLUS_EXPR;
+ /* The code can only be PLUS_EXPR or MINUS_EXPR. */
+ gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
+ tree_code incr_op = code;
edge pe = loop_preheader_edge (loop);
if (var != NULL_TREE)
@@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
tree ctr_before, ctr_after;
gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
exit_if = as_a <gcond *> (gsi_stmt (bsi));
- create_iv (exit_base, exit_step, NULL_TREE, loop,
+ create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
&bsi, false, &ctr_before, &ctr_after);
gimple_cond_set_code (exit_if, exit_cmp);
gimple_cond_set_lhs (exit_if, ctr_after);
@@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
gsi = gsi_last_bb (loop->latch);
else
gsi = gsi_last_nondebug_bb (loop->header);
- create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
- loop, &gsi, bump_in_latch, &var_before, NULL);
+ create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
+ NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
rewrite_all_phi_nodes_with_iv (loop, var_before);
diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
index d49273a3987..bda09f51d56 100644
--- a/gcc/tree-ssa-loop-manip.h
+++ b/gcc/tree-ssa-loop-manip.h
@@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
typedef void (*transform_callback)(class loop *, void *);
-extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
- bool, tree *, tree *);
+extern void create_iv (tree, tree_code, tree, tree, class loop *,
+ gimple_stmt_iterator *, bool, tree *, tree *);
extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
extern void verify_loop_closed_ssa (bool, class loop * = NULL);
diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index 6721ab6efc4..5c9103b16e5 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (aggr_ptr_init,
+ create_iv (aggr_ptr_init, PLUS_EXPR,
fold_convert (aggr_ptr_type, iv_step),
aggr_ptr, loop, &incr_gsi, insert_after,
&indx_before_incr, &indx_after_incr);
@@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
{
standard_iv_increment_position (containing_loop, &incr_gsi,
&insert_after);
- create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
- containing_loop, &incr_gsi, insert_after, &indx_before_incr,
- &indx_after_incr);
+ create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
+ aggr_ptr, containing_loop, &incr_gsi, insert_after,
+ &indx_before_incr, &indx_after_incr);
incr = gsi_stmt (incr_gsi);
/* Copy the points-to information if it exists. */
diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
index 44bd5f2c805..d63ded5d4f0 100644
--- a/gcc/tree-vect-loop-manip.cc
+++ b/gcc/tree-vect-loop-manip.cc
@@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
return false;
}
+/* Try to use permutes to define the lens in DEST_RGM using the lens
+ in SRC_RGM, given that the former has twice as many lens as the
+ latter. Return true on success, adding any new statements to SEQ. */
+
+static bool
+vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
+ rgroup_controls *dest_rgm,
+ rgroup_controls *src_rgm)
+{
+ tree ctrl_type = dest_rgm->type;
+ poly_uint64 nitems_per_ctrl
+ = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
+ if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
+ {
+ for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
+ {
+ tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
+ tree dest = dest_rgm->controls[i];
+ gassign *stmt;
+ if (i == 0)
+ {
+ /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
+ tree factor = build_int_cst (iv_type, nitems_per_ctrl);
+ stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
+ gimple_seq_add_stmt (seq, stmt);
+ }
+ else
+ {
+ /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
+ tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
+ tree temp = make_ssa_name (iv_type);
+ stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
+ gimple_seq_add_stmt (seq, stmt);
+ stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
+ gimple_seq_add_stmt (seq, stmt);
+ }
+ }
+ return true;
+ }
+ return false;
+}
+
/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
for all the rgroup controls in RGC and return a control that is nonzero
when the loop needs to iterate. Add any new preheader statements to
@@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
gimple_stmt_iterator incr_gsi;
bool insert_after;
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
- &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
+ create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
+ loop, &incr_gsi, insert_after, &index_before_incr,
+ &index_after_incr);
tree zero_index = build_int_cst (compare_type, 0);
tree test_index, test_limit, first_limit;
@@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
return next_ctrl;
}
+/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
+ for all the rgroup controls in RGC and return a control that is nonzero
+ when the loop needs to iterate. Add any new preheader statements to
+ PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
+
+ RGC belongs to loop LOOP. The loop originally iterated NITERS
+ times and has been vectorized according to LOOP_VINFO.
+
+ Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
+ to TEST_LIMIT - bias.
+
+ In vect_set_loop_controls_by_select_vl, we are iterating from start at
+ IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
+ IFN_SELECT_VL pattern.
+
+ 1. Single rgroup, the Gimple IR should be:
+
+ # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
+ # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
+ # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
+ # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
+ # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
+ _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
+ ivtmp_15 = _28 * 4;
+ vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
+ _1 = B[i_10];
+ .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
+ i_7 = i_10 + 1;
+ vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
+ vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
+ ivtmp_27 = ivtmp_26 - _28;
+ if (ivtmp_27 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 7>; [16.67%]
+
+ Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
+ data reference pointer IV.
+
+ 1). The result of .SELECT_VL:
+ _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
+ The _28 is not necessary to be VF in any iteration, instead, we allow
+ _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
+ pattern allows target have various flexible optimizations in vector
+ loop iterations. Target like RISC-V has special application vector
+ length calculation instruction which will distribute even workload
+ in the last 2 iterations.
+
+ Other example is that we can allow even generate _28 <= VF / 2 so
+ that some machine can run vector codes in low power mode.
+
+ 2). Loop control IV:
+ ivtmp_27 = ivtmp_26 - _28;
+ if (ivtmp_27 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 7>; [16.67%]
+
+ This is the saturating-subtraction towards zero, the outcome of
+ .SELECT_VL wil make ivtmp_27 never underflow zero.
+
+ 3). Data reference pointer IV:
+ ivtmp_15 = _28 * 4;
+ vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
+ vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
+
+ The pointer IV is adjusted accurately according to the .SELECT_VL.
+
+ 2. Multiple rgroup, the Gimple IR should be:
+
+ # i_23 = PHI <i_20(6), 0(11)>
+ # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
+ # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
+ # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
+ # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ _1 = i_23 * 2;
+ _2 = (long unsigned int) _1;
+ _3 = _2 * 2;
+ _4 = f_15(D) + _3;
+ _5 = _2 + 1;
+ _6 = _5 * 2;
+ _7 = f_15(D) + _6;
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ _8 = (long unsigned int) i_23;
+ _9 = _8 * 4;
+ _10 = d_18(D) + _9;
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+ i_20 = i_23 + 1;
+ vectp_f.8_52 = vectp_f.8_56 + 16;
+ vectp_d.10_60 = vectp_d.10_63 + 16;
+ ivtmp_74 = ivtmp_73 - _75;
+ ivtmp_71 = ivtmp_70 - _72;
+ if (ivtmp_74 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 13>; [16.67%]
+
+ Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
+ rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
+ iteration amount for mutiple rgroups.
+
+ The analysis of the flow of multiple rgroups:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ ...
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+
+ We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
+ to be processed in each iteration.
+
+ The related STOREs:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+ Since these 2 STOREs store 2 vectors that the second vector is half elements
+ of the first vector. So the length of second STORE will be _64 = _72 / 2;
+ It's similar to the VIEW_CONVERT of handling masks in SLP.
+
+ 3. Multiple rgroups for non-SLP auto-vectorization.
+
+ # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
+ # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
+ # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
+ _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+ _29 = (void *) ivtmp.35_10;
+ _7 = &MEM <vector([4,4]) int> [(int *)_29];
+ vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
+ _33 = _29 + POLY_INT_CST [16, 16];
+ _34 = &MEM <vector([4,4]) int> [(int *)_33];
+ vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
+ vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
+ _30 = (void *) ivtmp.36_2;
+ _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
+ .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
+ ivtmp_27 = ivtmp_26 - _28;
+ ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
+ ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
+ if (ivtmp_27 != 0)
+ goto <bb 4>; [83.33%]
+ else
+ goto <bb 5>; [16.67%]
+
+ The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+
+ The length of first half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+
+ The length of second half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+
+ 1). _28 always <= POLY_INT_CST [8, 8].
+ 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
+ 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
+*/
+
+static tree
+vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
+ gimple_seq *preheader_seq,
+ gimple_seq *header_seq,
+ rgroup_controls *rgc, tree niters,
+ unsigned int controls_length)
+{
+ tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
+ /* We are not allowing masked approach in SELECT_VL. */
+ gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
+
+ tree ctrl_type = rgc->type;
+ unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
+ poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
+ poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+
+ /* Calculate the maximum number of item values that the rgroup
+ handles in total, the number that it handles for each iteration
+ of the vector loop. */
+ tree nitems_total = niters;
+ if (nitems_per_iter != 1)
+ {
+ /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
+ these multiplications don't overflow. */
+ tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
+ nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
+ nitems_total, compare_factor);
+ }
+
+ /* Convert the comparison value to the IV type (either a no-op or
+ a promotion). */
+ nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
+
+ /* Create an induction variable that counts the number of items
+ processed. */
+ tree index_before_incr, index_after_incr;
+ gimple_stmt_iterator incr_gsi;
+ bool insert_after;
+ standard_iv_increment_position (loop, &incr_gsi, &insert_after);
+
+ /* Test the decremented IV, which will never underflow 0 since we have
+ IFN_SELECT_VL to gurantee that. */
+ tree test_limit = nitems_total;
+
+ /* Provide a definition of each control in the group. */
+ tree ctrl;
+ unsigned int i;
+ FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
+ {
+ /* Previous controls will cover BIAS items. This control covers the
+ next batch. */
+ poly_uint64 bias = nitems_per_ctrl * i;
+ tree bias_tree = build_int_cst (iv_type, bias);
+
+ /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
+ BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
+ control and adjust the bound down by BIAS. */
+ tree this_test_limit = test_limit;
+ if (i != 0)
+ {
+ this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
+ this_test_limit, bias_tree);
+ this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
+ this_test_limit, bias_tree);
+ }
+
+ /* Create decrement IV. */
+ create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
+ insert_after, &index_before_incr, &index_after_incr);
+
+ poly_uint64 final_vf = vf * nitems_per_iter;
+ tree vf_step = build_int_cst (iv_type, final_vf);
+ tree res_len;
+ if (nitems_per_iter != 1 || controls_length != 1)
+ {
+ /* For SLP, we can't allow non-VF number of elements to be processed
+ in non-final iteration. We force the number of elements to be
+ processed in each non-final iteration is VF elements. If we allow
+ non-VF elements processing in non-final iteration will make SLP too
+ complicated and produce inferior codegen.
+
+ For example:
+
+ If non-final iteration process VF elements.
+
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
+ ...
+
+ If non-final iteration process non-VF elements.
+
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
+ if (_71 % 2 == 0)
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
+ else
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
+ ...
+
+ This is the simple case of 2-elements interleaved vector SLP. We
+ consider other interleave vector, the situation will become more
+ complicated. */
+ res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
+ index_before_incr, vf_step);
+ }
+ else
+ {
+ res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
+ index_before_incr, vf_step);
+ }
+ gassign *assign = gimple_build_assign (ctrl, res_len);
+ gimple_seq_add_stmt (header_seq, assign);
+ }
+
+ return index_after_incr;
+}
+
/* Set up the iteration condition and rgroup controls for LOOP, given
that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
@@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
+ bool use_vl_p = !use_masks_p
+ && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
+ OPTIMIZE_FOR_SPEED);
unsigned int compare_precision = TYPE_PRECISION (compare_type);
tree orig_niters = niters;
@@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
continue;
}
+ if (use_vl_p && rgc->controls.length () != 1)
+ {
+ rgroup_controls *sub_rgc
+ = &(*controls)[nmasks / rgc->controls.length () - 1];
+ if (!sub_rgc->controls.is_empty ()
+ && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
+ sub_rgc))
+ continue;
+ }
+
/* See whether zero-based IV would ever generate all-false masks
or zero length before wrapping around. */
bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
/* Set up all controls for this group. */
- test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
- &preheader_seq,
- &header_seq,
- loop_cond_gsi, rgc,
- niters, niters_skip,
- might_wrap_p);
+ if (use_vl_p)
+ test_ctrl
+ = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
+ &preheader_seq, &header_seq,
+ rgc, niters, controls->length ());
+ else
+ test_ctrl
+ = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
+ &header_seq, loop_cond_gsi, rgc,
+ niters, niters_skip,
+ might_wrap_p);
}
/* Emit all accumulated statements. */
@@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
}
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (init, step, NULL_TREE, loop,
+ create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
&incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
true, NULL_TREE, true,
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index 6ea0f21fd13..d7cf92576d1 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
gimple_stmt_iterator incr_gsi;
bool insert_after;
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
- create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
+ create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
insert_after, &indx_before_incr, &indx_after_incr);
/* Next create a new phi node vector (NEW_PHI_TREE) which starts
@@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
}
/* Given a complete set of length LENS, extract length number INDEX for an
- rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
+ rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
+ Insert any set-up statements before GSI. */
tree
-vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
- unsigned int nvectors, unsigned int index)
+vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
+ vec_loop_lens *lens, unsigned int nvectors, tree vectype,
+ unsigned int index)
{
rgroup_controls *rgl = &(*lens)[nvectors - 1];
bool use_bias_adjusted_len =
LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
/* Populate the rgroup's len array, if this is the first time we've
used it. */
@@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
if (use_bias_adjusted_len)
return rgl->bias_adjusted_ctrl;
+ else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
+ OPTIMIZE_FOR_SPEED))
+ {
+ tree loop_len = rgl->controls[index];
+ poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
+ poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
+ if (maybe_ne (nunits1, nunits2))
+ {
+ /* A loop len for data type X can be reused for data type Y
+ if X has N times more elements than Y and if Y's elements
+ are N times bigger than X's. */
+ gcc_assert (multiple_p (nunits1, nunits2));
+ unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
+ gimple_seq seq = NULL;
+ loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
+ build_int_cst (iv_type, factor));
+ if (seq)
+ gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
+ }
+ return loop_len;
+ }
else
return rgl->controls[index];
}
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 3ad6a7d28d7..26813ab56e9 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
return iv_step;
}
+/* Prepare the pointer IVs which needs to be updated by a variable amount.
+ Such variable amount is the outcome of .SELECT_VL. In this case, we can
+ allow each iteration process the flexible number of elements as long as
+ the number <= vf elments.
+
+ Return data reference according to SELECT_VL.
+ If new statements are needed, insert them before GSI. */
+
+static tree
+get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
+ tree aggr_type, class loop *at_loop, tree offset,
+ tree *dummy, gimple_stmt_iterator *gsi,
+ bool simd_lane_access_p, vec_loop_lens *loop_lens,
+ dr_vec_info *dr_info,
+ vect_memory_access_type memory_access_type)
+{
+ loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
+ tree step = vect_dr_behavior (vinfo, dr_info)->step;
+
+ /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
+ IVs are updated by variable amount but we will support them in the future.
+ */
+ gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
+ && memory_access_type != VMAT_LOAD_STORE_LANES);
+
+ /* When we support SELECT_VL pattern, we dynamic adjust
+ the memory address by .SELECT_VL result.
+
+ The result of .SELECT_VL is the number of elements to
+ be processed of each iteration. So the memory address
+ adjustment operation should be:
+
+ bytesize = GET_MODE_SIZE (element_mode (aggr_type));
+ addr = addr + .SELECT_VL (ARG..) * bytesize;
+ */
+ gimple *ptr_incr;
+ tree loop_len
+ = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
+ tree len_type = TREE_TYPE (loop_len);
+ poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
+ /* Since the outcome of .SELECT_VL is element size, we should adjust
+ it into bytesize so that it can be used in address pointer variable
+ amount IVs adjustment. */
+ tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
+ build_int_cst (len_type, bytesize));
+ if (tree_int_cst_sgn (step) == -1)
+ tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
+ tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
+ gassign *assign = gimple_build_assign (bump, tmp);
+ gsi_insert_before (gsi, assign, GSI_SAME_STMT);
+ return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
+ dummy, gsi, &ptr_incr, simd_lane_access_p,
+ bump);
+}
+
/* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
static bool
@@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
- create_iv (stride_base, ivstep, NULL,
+ create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
loop, &incr_gsi, insert_after,
&offvar, NULL);
incr = gsi_stmt (incr_gsi);
@@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
slp_node, &gs_info, &dataref_ptr,
&vec_offsets);
+ else if (loop_lens && loop_lens->length () == 1
+ && direct_internal_fn_supported_p (
+ IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
+ OPTIMIZE_FOR_SPEED)
+ && memory_access_type != VMAT_INVARIANT)
+ dataref_ptr
+ = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
+ simd_lane_access_p ? loop : NULL,
+ offset, &dummy, gsi,
+ simd_lane_access_p, loop_lens,
+ dr_info, memory_access_type);
else
dataref_ptr
= vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
@@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
else if (loop_lens)
{
tree final_len
- = vect_get_loop_len (loop_vinfo, loop_lens,
- vec_num * ncopies, vec_num * j + i);
+ = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ vec_num * ncopies, vectype,
+ vec_num * j + i);
tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
machine_mode vmode = TYPE_MODE (vectype);
opt_machine_mode new_ovmode
@@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
- create_iv (stride_base, ivstep, NULL,
+ create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
loop, &incr_gsi, insert_after,
&offvar, NULL);
@@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
slp_node, &gs_info, &dataref_ptr,
&vec_offsets);
}
+ else if (loop_lens && loop_lens->length () == 1
+ && direct_internal_fn_supported_p (
+ IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
+ OPTIMIZE_FOR_SPEED)
+ && memory_access_type != VMAT_INVARIANT)
+ dataref_ptr
+ = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
+ at_loop, offset, &dummy, gsi,
+ simd_lane_access_p, loop_lens,
+ dr_info, memory_access_type);
else
dataref_ptr
= vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
@@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
else if (loop_lens && memory_access_type != VMAT_INVARIANT)
{
tree final_len
- = vect_get_loop_len (loop_vinfo, loop_lens,
- vec_num * ncopies,
+ = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
+ vec_num * ncopies, vectype,
vec_num * j + i);
tree ptr = build_int_cst (ref_type,
align * BITS_PER_UNIT);
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 9cf2fb23fe3..357bc5c7315 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
unsigned int, tree, unsigned int);
extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
tree, unsigned int);
-extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
- unsigned int);
+extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
+ unsigned int, tree, unsigned int);
extern gimple_seq vect_gen_len (tree, tree, tree, tree);
extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
--
2.36.1
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-04 13:25 [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support juzhe.zhong
2023-05-05 23:41 ` 钟居哲
@ 2023-05-07 15:19 ` Jeff Law
2023-05-07 21:54 ` 钟居哲
2023-05-09 10:52 ` juzhe.zhong
2023-05-08 5:35 ` Kewen.Lin
2023-05-10 16:45 ` Richard Sandiford
3 siblings, 2 replies; 25+ messages in thread
From: Jeff Law @ 2023-05-07 15:19 UTC (permalink / raw)
To: juzhe.zhong, gcc-patches; +Cc: richard.sandiford, rguenther
On 5/4/23 07:25, juzhe.zhong@rivai.ai wrote:
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
So at a high level this is pretty good. I think there's some
improvements we should make in the documentation and comments, but I'm
comfortable with most of the implementation details.
>
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
This reads rather poorly. Is this still accurate?
Set operand 0 to the number of active elements in a vector to be updated
in a loop iteration based on the total number of elements to be updated,
the vectorization factor and vector properties of the target.
> +operand 1 is the total elements need to be updated value.
operand 1 is the total elements in the vector to be updated.
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
This allows dynamic adjustment of the number of elements processed each
loop iteration. -- is that still accurate and does it read better?
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
It's been pretty standard to stick with just PLUS_EXPR for this stuff
and instead negate the constant to produce the same effect as
MINUS_EXPR. Is there a reason we're not continuing that practice?
Sorry if you've answered this already -- if you have, you can just point
me at the prior discussion and I'll read it.
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
I would suggest not using "permute" in this description. When I read
permute in the context of vectorization, I think of a vector permute to
scramble elements within a vector.
This looks like you're just adjusting how many vector elements you're
operating on.
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
Looks like you may have mixed up spaces and tabs in the above comment.
Just a nit, but let's go ahead and get it fixed.
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
>
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
When you break a line with a logical like this, go ahead and add
parenthesis and make sure the logical aligns just after the paren. ie
bool use_vl_p = (!use_masks_p
&& direct....
Alternately, compute the direct_itnernal_fn_supported_p into its own
boolean and then you don't need as much line wrapping.
In general, don't be afraid to use extra temporaries if doing so
improves readability.
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
This looks like a good example of code that would be easier to read if
the call to direct_internal-fn_supported_p was saved into a temporary.
Similarly for the instance you added in vectorizable_load.
I'd like to get this patch wrapped up soon. But I also want to give
both Richards a chance to chime in with their concerns.
Thanks,
Jeff
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-07 15:19 ` Jeff Law
@ 2023-05-07 21:54 ` 钟居哲
2023-05-09 10:52 ` juzhe.zhong
1 sibling, 0 replies; 25+ messages in thread
From: 钟居哲 @ 2023-05-07 21:54 UTC (permalink / raw)
To: Jeff Law, gcc-patches; +Cc: richard.sandiford, rguenther
[-- Attachment #1: Type: text/plain, Size: 10389 bytes --]
>> It's been pretty standard to stick with just PLUS_EXPR for this stuff
>> and instead negate the constant to produce the same effect as
>> MINUS_EXPR. Is there a reason we're not continuing that practice?
>> Sorry if you've answered this already -- if you have, you can just point
>> me at the prior discussion and I'll read it.
Richard (Sandiford) has answered this question:
https://gcc.gnu.org/pipermail/gcc-patches/2023-April/616745.html
And as Richard (Biener) said, it will make IVOPTs failed if we have variable IVs.
However, we already have implemented variable IVs in downstream RVV GCC
and works fine and I don't see any bad codegen so far. So, I think it may not
be a serious issue for RVV.
Besides, this implementation is not my idea which is just following the guide coming
from RVV ISA spec. And also inspired by the implementation coming from LLVM.
Reference:
1). RVV ISA: https://github.com/riscv/riscv-v-spec/blob/master/example/vvaddint32.s
2). LLVM length stuff implementation (Should note that "get_vector_length" pattern in LLVM is
totally doing the same thing as "select_vl" pattern in GCC here:
https://reviews.llvm.org/D99750
vvaddint32:
vsetvli t0, a0, e32, ta, ma # Set vector length based on 32-bit vectors
vle32.v v0, (a1) # Get first vector
sub a0, a0, t0 # Decrement number done
slli t0, t0, 2 # Multiply number done by 4 bytes
add a1, a1, t0 # Bump pointer
vle32.v v1, (a2) # Get second vector
add a2, a2, t0 # Bump pointer
vadd.vv v2, v0, v1 # Sum vectors
vse32.v v2, (a3) # Store result
add a3, a3, t0 # Bump pointer
bnez a0, vvaddint32 # Loop back
ret # Finished
Notice "sub a0, a0, t0", the "t0" is the variable coming from "vsetvli t0, a0, e32, ta, ma"
which is generated by "get_vector_length" in LLVM, or similiar it is also generatd by "select_vl" in GCC too.
Other comments from Jeff will be addressed in the next patch (V5), I will wait for
Richards (both Sandiford && Biener that are the experts in Loop Vectorizer) comments.
Then send V5 patch which is including all comments from Jeff && Richards (Sandiford && Biener).
Thanks.
juzhe.zhong@rivai.ai
From: Jeff Law
Date: 2023-05-07 23:19
To: juzhe.zhong; gcc-patches
CC: richard.sandiford; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
On 5/4/23 07:25, juzhe.zhong@rivai.ai wrote:
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
So at a high level this is pretty good. I think there's some
improvements we should make in the documentation and comments, but I'm
comfortable with most of the implementation details.
>
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
This reads rather poorly. Is this still accurate?
Set operand 0 to the number of active elements in a vector to be updated
in a loop iteration based on the total number of elements to be updated,
the vectorization factor and vector properties of the target.
> +operand 1 is the total elements need to be updated value.
operand 1 is the total elements in the vector to be updated.
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
This allows dynamic adjustment of the number of elements processed each
loop iteration. -- is that still accurate and does it read better?
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
It's been pretty standard to stick with just PLUS_EXPR for this stuff
and instead negate the constant to produce the same effect as
MINUS_EXPR. Is there a reason we're not continuing that practice?
Sorry if you've answered this already -- if you have, you can just point
me at the prior discussion and I'll read it.
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
I would suggest not using "permute" in this description. When I read
permute in the context of vectorization, I think of a vector permute to
scramble elements within a vector.
This looks like you're just adjusting how many vector elements you're
operating on.
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
Looks like you may have mixed up spaces and tabs in the above comment.
Just a nit, but let's go ahead and get it fixed.
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
>
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
When you break a line with a logical like this, go ahead and add
parenthesis and make sure the logical aligns just after the paren. ie
bool use_vl_p = (!use_masks_p
&& direct....
Alternately, compute the direct_itnernal_fn_supported_p into its own
boolean and then you don't need as much line wrapping.
In general, don't be afraid to use extra temporaries if doing so
improves readability.
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
This looks like a good example of code that would be easier to read if
the call to direct_internal-fn_supported_p was saved into a temporary.
Similarly for the instance you added in vectorizable_load.
I'd like to get this patch wrapped up soon. But I also want to give
both Richards a chance to chime in with their concerns.
Thanks,
Jeff
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-04 13:25 [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support juzhe.zhong
2023-05-05 23:41 ` 钟居哲
2023-05-07 15:19 ` Jeff Law
@ 2023-05-08 5:35 ` Kewen.Lin
2023-05-08 6:27 ` juzhe.zhong
2023-05-10 16:45 ` Richard Sandiford
3 siblings, 1 reply; 25+ messages in thread
From: Kewen.Lin @ 2023-05-08 5:35 UTC (permalink / raw)
To: juzhe.zhong, rguenther, richard.sandiford; +Cc: gcc-patches
Hi,
on 2023/5/4 21:25, juzhe.zhong@rivai.ai wrote:
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
> Thanks.
>
> ---
> gcc/cfgloopmanip.cc | 2 +-
> gcc/doc/md.texi | 34 +++
> gcc/gimple-loop-interchange.cc | 2 +-
> gcc/internal-fn.def | 1 +
> gcc/optabs.def | 1 +
> gcc/tree-ssa-loop-ivcanon.cc | 2 +-
> gcc/tree-ssa-loop-ivopts.cc | 2 +-
> gcc/tree-ssa-loop-manip.cc | 18 +-
> gcc/tree-ssa-loop-manip.h | 4 +-
> gcc/tree-vect-data-refs.cc | 8 +-
> gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
> gcc/tree-vect-loop.cc | 32 ++-
> gcc/tree-vect-stmts.cc | 89 +++++++-
> gcc/tree-vectorizer.h | 4 +-
> 14 files changed, 535 insertions(+), 38 deletions(-)
>
> diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
> index 0e3ad8ed742..6e09dcbb0b1 100644
> --- a/gcc/cfgloopmanip.cc
> +++ b/gcc/cfgloopmanip.cc
> @@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
> }
>
> gsi = gsi_last_bb (loop_header);
> - create_iv (initial_value, stride, iv, loop, &gsi, false,
> + create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
> iv_before, iv_after);
>
> /* Insert loop exit condition. */
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
> index 1b77bfd46b2..e5590374e59 100644
> --- a/gcc/gimple-loop-interchange.cc
> +++ b/gcc/gimple-loop-interchange.cc
> @@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
> tree var_before, var_after;
> tree base = unshare_expr (iv->init_expr);
> tree step = unshare_expr (iv->step);
> - create_iv (base, step, SSA_NAME_VAR (iv->var),
> + create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
> tgt.m_loop, &incr_pos, false, &var_before, &var_after);
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 7fe742c2ae7..6f6fa7d37f9 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
> DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
>
> DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
> +DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
> DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
> check_raw_ptrs, check_ptrs)
> DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 695f5911b30..b637471b76e 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
> OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
> OPTAB_D (len_load_optab, "len_load_$a")
> OPTAB_D (len_store_optab, "len_store_$a")
> +OPTAB_D (select_vl_optab, "select_vl$a")
> diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
> index e41ec73a52a..d4e113c44df 100644
> --- a/gcc/tree-ssa-loop-ivcanon.cc
> +++ b/gcc/tree-ssa-loop-ivcanon.cc
> @@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
> niter,
> build_int_cst (type, 1));
> incr_at = gsi_last_bb (in->src);
> - create_iv (niter,
> + create_iv (niter, PLUS_EXPR,
> build_int_cst (type, -1),
> NULL_TREE, loop,
> &incr_at, false, var_before, &var);
> diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
> index 78e8cbc75b5..8640fe2d487 100644
> --- a/gcc/tree-ssa-loop-ivopts.cc
> +++ b/gcc/tree-ssa-loop-ivopts.cc
> @@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
>
> base = unshare_expr (cand->iv->base);
>
> - create_iv (base, unshare_expr (cand->iv->step),
> + create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
> cand->var_before, data->current_loop,
> &incr_pos, after, &cand->var_before, &cand->var_after);
> }
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
>
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
> edge pe = loop_preheader_edge (loop);
>
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
>
> rewrite_all_phi_nodes_with_iv (loop, var_before);
>
> diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
> index d49273a3987..bda09f51d56 100644
> --- a/gcc/tree-ssa-loop-manip.h
> +++ b/gcc/tree-ssa-loop-manip.h
> @@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
>
> typedef void (*transform_callback)(class loop *, void *);
>
> -extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
> - bool, tree *, tree *);
> +extern void create_iv (tree, tree_code, tree, tree, class loop *,
> + gimple_stmt_iterator *, bool, tree *, tree *);
> extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
> extern void verify_loop_closed_ssa (bool, class loop * = NULL);
>
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index 6721ab6efc4..5c9103b16e5 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
>
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
>
> - create_iv (aggr_ptr_init,
> + create_iv (aggr_ptr_init, PLUS_EXPR,
> fold_convert (aggr_ptr_type, iv_step),
> aggr_ptr, loop, &incr_gsi, insert_after,
> &indx_before_incr, &indx_after_incr);
> @@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> {
> standard_iv_increment_position (containing_loop, &incr_gsi,
> &insert_after);
> - create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
> - containing_loop, &incr_gsi, insert_after, &indx_before_incr,
> - &indx_after_incr);
> + create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
> + aggr_ptr, containing_loop, &incr_gsi, insert_after,
> + &indx_before_incr, &indx_after_incr);
> incr = gsi_stmt (incr_gsi);
>
> /* Copy the points-to information if it exists. */
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> @@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
> - &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
> + create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
> + loop, &incr_gsi, insert_after, &index_before_incr,
> + &index_after_incr);
>
> tree zero_index = build_int_cst (compare_type, 0);
> tree test_index, test_limit, first_limit;
> @@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> return next_ctrl;
> }
>
> +/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> + for all the rgroup controls in RGC and return a control that is nonzero
> + when the loop needs to iterate. Add any new preheader statements to
> + PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
> +
> + RGC belongs to loop LOOP. The loop originally iterated NITERS
> + times and has been vectorized according to LOOP_VINFO.
> +
> + Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
> + to TEST_LIMIT - bias.
> +
> + In vect_set_loop_controls_by_select_vl, we are iterating from start at
> + IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
> + IFN_SELECT_VL pattern.
> +
> + 1. Single rgroup, the Gimple IR should be:
> +
> + # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
> + # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
> + # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
> + # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
> + # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + ivtmp_15 = _28 * 4;
> + vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
> + _1 = B[i_10];
> + .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
> + i_7 = i_10 + 1;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
> + data reference pointer IV.
> +
> + 1). The result of .SELECT_VL:
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + The _28 is not necessary to be VF in any iteration, instead, we allow
> + _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
> + pattern allows target have various flexible optimizations in vector
> + loop iterations. Target like RISC-V has special application vector
> + length calculation instruction which will distribute even workload
> + in the last 2 iterations.
> +
> + Other example is that we can allow even generate _28 <= VF / 2 so
> + that some machine can run vector codes in low power mode.
> +
> + 2). Loop control IV:
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + This is the saturating-subtraction towards zero, the outcome of
> + .SELECT_VL wil make ivtmp_27 never underflow zero.
> +
> + 3). Data reference pointer IV:
> + ivtmp_15 = _28 * 4;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> +
> + The pointer IV is adjusted accurately according to the .SELECT_VL.
> +
> + 2. Multiple rgroup, the Gimple IR should be:
> +
> + # i_23 = PHI <i_20(6), 0(11)>
> + # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
> + # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
> + # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
> + # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + _1 = i_23 * 2;
> + _2 = (long unsigned int) _1;
> + _3 = _2 * 2;
> + _4 = f_15(D) + _3;
> + _5 = _2 + 1;
> + _6 = _5 * 2;
> + _7 = f_15(D) + _6;
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _8 = (long unsigned int) i_23;
> + _9 = _8 * 4;
> + _10 = d_18(D) + _9;
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + i_20 = i_23 + 1;
> + vectp_f.8_52 = vectp_f.8_56 + 16;
> + vectp_d.10_60 = vectp_d.10_63 + 16;
> + ivtmp_74 = ivtmp_73 - _75;
> + ivtmp_71 = ivtmp_70 - _72;
> + if (ivtmp_74 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 13>; [16.67%]
> +
> + Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
> + rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
> + iteration amount for mutiple rgroups.
> +
> + The analysis of the flow of multiple rgroups:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + ...
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> +
> + We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
> + to be processed in each iteration.
> +
> + The related STOREs:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + Since these 2 STOREs store 2 vectors that the second vector is half elements
> + of the first vector. So the length of second STORE will be _64 = _72 / 2;
> + It's similar to the VIEW_CONVERT of handling masks in SLP.
> +
> + 3. Multiple rgroups for non-SLP auto-vectorization.
> +
> + # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
> + # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
> + # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
> + _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> + _29 = (void *) ivtmp.35_10;
> + _7 = &MEM <vector([4,4]) int> [(int *)_29];
> + vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
> + _33 = _29 + POLY_INT_CST [16, 16];
> + _34 = &MEM <vector([4,4]) int> [(int *)_33];
> + vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
> + vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
> + _30 = (void *) ivtmp.36_2;
> + _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
> + .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
> + ivtmp_27 = ivtmp_26 - _28;
> + ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
> + ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
> + if (ivtmp_27 != 0)
> + goto <bb 4>; [83.33%]
> + else
> + goto <bb 5>; [16.67%]
> +
> + The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> +
> + The length of first half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> +
> + The length of second half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> +
> + 1). _28 always <= POLY_INT_CST [8, 8].
> + 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
> + 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
> +*/
> +
> +static tree
> +vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
> + gimple_seq *preheader_seq,
> + gimple_seq *header_seq,
> + rgroup_controls *rgc, tree niters,
> + unsigned int controls_length)
> +{
> + tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + /* We are not allowing masked approach in SELECT_VL. */
> + gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
> +
> + tree ctrl_type = rgc->type;
> + unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
> + poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
> + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
> +
> + /* Calculate the maximum number of item values that the rgroup
> + handles in total, the number that it handles for each iteration
> + of the vector loop. */
> + tree nitems_total = niters;
> + if (nitems_per_iter != 1)
> + {
> + /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
> + these multiplications don't overflow. */
> + tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
> + nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
> + nitems_total, compare_factor);
> + }
> +
> + /* Convert the comparison value to the IV type (either a no-op or
> + a promotion). */
> + nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
> +
> + /* Create an induction variable that counts the number of items
> + processed. */
> + tree index_before_incr, index_after_incr;
> + gimple_stmt_iterator incr_gsi;
> + bool insert_after;
> + standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> +
> + /* Test the decremented IV, which will never underflow 0 since we have
> + IFN_SELECT_VL to gurantee that. */
> + tree test_limit = nitems_total;
> +
> + /* Provide a definition of each control in the group. */
> + tree ctrl;
> + unsigned int i;
> + FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
> + {
> + /* Previous controls will cover BIAS items. This control covers the
> + next batch. */
> + poly_uint64 bias = nitems_per_ctrl * i;
> + tree bias_tree = build_int_cst (iv_type, bias);
> +
> + /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
> + BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
> + control and adjust the bound down by BIAS. */
> + tree this_test_limit = test_limit;
> + if (i != 0)
> + {
> + this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + }
> +
> + /* Create decrement IV. */
> + create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
> + insert_after, &index_before_incr, &index_after_incr);
> +
> + poly_uint64 final_vf = vf * nitems_per_iter;
> + tree vf_step = build_int_cst (iv_type, final_vf);
> + tree res_len;
> + if (nitems_per_iter != 1 || controls_length != 1)
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
> +
> + For example:
> +
> + If non-final iteration process VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + ...
> +
> + If non-final iteration process non-VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + if (_71 % 2 == 0)
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + else
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
> + ...
> +
> + This is the simple case of 2-elements interleaved vector SLP. We
> + consider other interleave vector, the situation will become more
> + complicated. */
> + res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
> + index_before_incr, vf_step);
> + }
> + else
> + {
> + res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
> + index_before_incr, vf_step);
> + }
> + gassign *assign = gimple_build_assign (ctrl, res_len);
> + gimple_seq_add_stmt (header_seq, assign);
> + }
> +
> + return index_after_incr;
> +}
> +
> /* Set up the iteration condition and rgroup controls for LOOP, given
> that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
> loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
>
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
> unsigned int compare_precision = TYPE_PRECISION (compare_type);
> tree orig_niters = niters;
>
> @@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> continue;
> }
>
> + if (use_vl_p && rgc->controls.length () != 1)
> + {
> + rgroup_controls *sub_rgc
> + = &(*controls)[nmasks / rgc->controls.length () - 1];
> + if (!sub_rgc->controls.is_empty ()
> + && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
> + sub_rgc))
> + continue;
> + }
> +
> /* See whether zero-based IV would ever generate all-false masks
> or zero length before wrapping around. */
> bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
>
> /* Set up all controls for this group. */
> - test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
> - &preheader_seq,
> - &header_seq,
> - loop_cond_gsi, rgc,
> - niters, niters_skip,
> - might_wrap_p);
> + if (use_vl_p)
> + test_ctrl
> + = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
> + &preheader_seq, &header_seq,
> + rgc, niters, controls->length ());
> + else
> + test_ctrl
> + = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
> + &header_seq, loop_cond_gsi, rgc,
> + niters, niters_skip,
> + might_wrap_p);
Sorry for chiming in, I had some concern here.
We already have some handlings for partial vector in length in vect_set_loop_controls_directly
(actually it deals with both mask and length), the proposed vect_set_loop_controls_by_select_vl
for select_vl looks like a variant for partial vector in length (comparing to the normal MIN),
and also adopts decrement IV. IMHO, it seems better to modify/extend the current handling in
vect_set_loop_controls_directly for length, or factor out the handlings for length there and
extend the factored one. Otherwise, it means we have two sets of handlings for partial vector
in lengths, it looks bad to maintain. As the previous discussion, adopting decrement IV is an
enhancement for loop control, it's good for both cases w/ or w/o select_vl. If the above
understanding is correct, the possible steps seem to be:
- factor out the handling for length (* optional)
- modify it with decrement IV
- extend it with select_vl.
In future if some RVV vendor wants to degenerate select_vl to min, it can just adopt the same
handlings with min by not defining select_vl optab.
BR,
Kewen
> }
>
> /* Emit all accumulated statements. */
> @@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
> }
>
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (init, step, NULL_TREE, loop,
> + create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
> &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
> indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
> true, NULL_TREE, true,
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 6ea0f21fd13..d7cf92576d1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
> + create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
> insert_after, &indx_before_incr, &indx_after_incr);
>
> /* Next create a new phi node vector (NEW_PHI_TREE) which starts
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
>
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
>
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
>
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
>
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
>
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
>
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
>
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-08 5:35 ` Kewen.Lin
@ 2023-05-08 6:27 ` juzhe.zhong
2023-05-08 7:55 ` Kewen.Lin
0 siblings, 1 reply; 25+ messages in thread
From: juzhe.zhong @ 2023-05-08 6:27 UTC (permalink / raw)
To: linkw, rguenther, richard.sandiford; +Cc: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 43521 bytes --]
Hi, Kewen.>> Sorry for chiming in, I had some concern here.
>> We already have some handlings for partial vector in length in vect_set_loop_controls_directly
>>(actually it deals with both mask and length), the proposed vect_set_loop_controls_by_select_vl
>>for select_vl looks like a variant for partial vector in length (comparing to the normal MIN),
>>and also adopts decrement IV. IMHO, it seems better to modify/extend the current handling in
>>vect_set_loop_controls_directly for length, or factor out the handlings for length there and
>>extend the factored one. Otherwise, it means we have two sets of handlings for partial vector
>>in lengths, it looks bad to maintain. As the previous discussion, adopting decrement IV is an
>>enhancement for loop control, it's good for both cases w/ or w/o select_vl. If the above
>>understanding is correct, the possible steps seem to be:
>> - factor out the handling for length (* optional)
>> - modify it with decrement IV
>> - extend it with select_vl.
>>In future if some RVV vendor wants to degenerate select_vl to min, it can just adopt the same
>>handlings with min by not defining select_vl optab.
You mean like this:
doing this inside vect_set_loop_controls_directly ?
if (use_while_len_p)
return vect_set_loop_controls_by_while_len (...)
Thanks.
juzhe.zhong@rivai.ai
From: Kewen.Lin
Date: 2023-05-08 13:35
To: juzhe.zhong; rguenther; richard.sandiford
CC: gcc-patches
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
Hi,
on 2023/5/4 21:25, juzhe.zhong@rivai.ai wrote:
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
> Thanks.
>
> ---
> gcc/cfgloopmanip.cc | 2 +-
> gcc/doc/md.texi | 34 +++
> gcc/gimple-loop-interchange.cc | 2 +-
> gcc/internal-fn.def | 1 +
> gcc/optabs.def | 1 +
> gcc/tree-ssa-loop-ivcanon.cc | 2 +-
> gcc/tree-ssa-loop-ivopts.cc | 2 +-
> gcc/tree-ssa-loop-manip.cc | 18 +-
> gcc/tree-ssa-loop-manip.h | 4 +-
> gcc/tree-vect-data-refs.cc | 8 +-
> gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
> gcc/tree-vect-loop.cc | 32 ++-
> gcc/tree-vect-stmts.cc | 89 +++++++-
> gcc/tree-vectorizer.h | 4 +-
> 14 files changed, 535 insertions(+), 38 deletions(-)
>
> diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
> index 0e3ad8ed742..6e09dcbb0b1 100644
> --- a/gcc/cfgloopmanip.cc
> +++ b/gcc/cfgloopmanip.cc
> @@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
> }
>
> gsi = gsi_last_bb (loop_header);
> - create_iv (initial_value, stride, iv, loop, &gsi, false,
> + create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
> iv_before, iv_after);
>
> /* Insert loop exit condition. */
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
> index 1b77bfd46b2..e5590374e59 100644
> --- a/gcc/gimple-loop-interchange.cc
> +++ b/gcc/gimple-loop-interchange.cc
> @@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
> tree var_before, var_after;
> tree base = unshare_expr (iv->init_expr);
> tree step = unshare_expr (iv->step);
> - create_iv (base, step, SSA_NAME_VAR (iv->var),
> + create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
> tgt.m_loop, &incr_pos, false, &var_before, &var_after);
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 7fe742c2ae7..6f6fa7d37f9 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
> DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
>
> DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
> +DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
> DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
> check_raw_ptrs, check_ptrs)
> DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 695f5911b30..b637471b76e 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
> OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
> OPTAB_D (len_load_optab, "len_load_$a")
> OPTAB_D (len_store_optab, "len_store_$a")
> +OPTAB_D (select_vl_optab, "select_vl$a")
> diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
> index e41ec73a52a..d4e113c44df 100644
> --- a/gcc/tree-ssa-loop-ivcanon.cc
> +++ b/gcc/tree-ssa-loop-ivcanon.cc
> @@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
> niter,
> build_int_cst (type, 1));
> incr_at = gsi_last_bb (in->src);
> - create_iv (niter,
> + create_iv (niter, PLUS_EXPR,
> build_int_cst (type, -1),
> NULL_TREE, loop,
> &incr_at, false, var_before, &var);
> diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
> index 78e8cbc75b5..8640fe2d487 100644
> --- a/gcc/tree-ssa-loop-ivopts.cc
> +++ b/gcc/tree-ssa-loop-ivopts.cc
> @@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
>
> base = unshare_expr (cand->iv->base);
>
> - create_iv (base, unshare_expr (cand->iv->step),
> + create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
> cand->var_before, data->current_loop,
> &incr_pos, after, &cand->var_before, &cand->var_after);
> }
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
>
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
> edge pe = loop_preheader_edge (loop);
>
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
>
> rewrite_all_phi_nodes_with_iv (loop, var_before);
>
> diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
> index d49273a3987..bda09f51d56 100644
> --- a/gcc/tree-ssa-loop-manip.h
> +++ b/gcc/tree-ssa-loop-manip.h
> @@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
>
> typedef void (*transform_callback)(class loop *, void *);
>
> -extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
> - bool, tree *, tree *);
> +extern void create_iv (tree, tree_code, tree, tree, class loop *,
> + gimple_stmt_iterator *, bool, tree *, tree *);
> extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
> extern void verify_loop_closed_ssa (bool, class loop * = NULL);
>
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index 6721ab6efc4..5c9103b16e5 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
>
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
>
> - create_iv (aggr_ptr_init,
> + create_iv (aggr_ptr_init, PLUS_EXPR,
> fold_convert (aggr_ptr_type, iv_step),
> aggr_ptr, loop, &incr_gsi, insert_after,
> &indx_before_incr, &indx_after_incr);
> @@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> {
> standard_iv_increment_position (containing_loop, &incr_gsi,
> &insert_after);
> - create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
> - containing_loop, &incr_gsi, insert_after, &indx_before_incr,
> - &indx_after_incr);
> + create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
> + aggr_ptr, containing_loop, &incr_gsi, insert_after,
> + &indx_before_incr, &indx_after_incr);
> incr = gsi_stmt (incr_gsi);
>
> /* Copy the points-to information if it exists. */
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> @@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
> - &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
> + create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
> + loop, &incr_gsi, insert_after, &index_before_incr,
> + &index_after_incr);
>
> tree zero_index = build_int_cst (compare_type, 0);
> tree test_index, test_limit, first_limit;
> @@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> return next_ctrl;
> }
>
> +/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> + for all the rgroup controls in RGC and return a control that is nonzero
> + when the loop needs to iterate. Add any new preheader statements to
> + PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
> +
> + RGC belongs to loop LOOP. The loop originally iterated NITERS
> + times and has been vectorized according to LOOP_VINFO.
> +
> + Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
> + to TEST_LIMIT - bias.
> +
> + In vect_set_loop_controls_by_select_vl, we are iterating from start at
> + IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
> + IFN_SELECT_VL pattern.
> +
> + 1. Single rgroup, the Gimple IR should be:
> +
> + # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
> + # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
> + # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
> + # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
> + # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + ivtmp_15 = _28 * 4;
> + vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
> + _1 = B[i_10];
> + .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
> + i_7 = i_10 + 1;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
> + data reference pointer IV.
> +
> + 1). The result of .SELECT_VL:
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + The _28 is not necessary to be VF in any iteration, instead, we allow
> + _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
> + pattern allows target have various flexible optimizations in vector
> + loop iterations. Target like RISC-V has special application vector
> + length calculation instruction which will distribute even workload
> + in the last 2 iterations.
> +
> + Other example is that we can allow even generate _28 <= VF / 2 so
> + that some machine can run vector codes in low power mode.
> +
> + 2). Loop control IV:
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + This is the saturating-subtraction towards zero, the outcome of
> + .SELECT_VL wil make ivtmp_27 never underflow zero.
> +
> + 3). Data reference pointer IV:
> + ivtmp_15 = _28 * 4;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> +
> + The pointer IV is adjusted accurately according to the .SELECT_VL.
> +
> + 2. Multiple rgroup, the Gimple IR should be:
> +
> + # i_23 = PHI <i_20(6), 0(11)>
> + # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
> + # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
> + # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
> + # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + _1 = i_23 * 2;
> + _2 = (long unsigned int) _1;
> + _3 = _2 * 2;
> + _4 = f_15(D) + _3;
> + _5 = _2 + 1;
> + _6 = _5 * 2;
> + _7 = f_15(D) + _6;
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _8 = (long unsigned int) i_23;
> + _9 = _8 * 4;
> + _10 = d_18(D) + _9;
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + i_20 = i_23 + 1;
> + vectp_f.8_52 = vectp_f.8_56 + 16;
> + vectp_d.10_60 = vectp_d.10_63 + 16;
> + ivtmp_74 = ivtmp_73 - _75;
> + ivtmp_71 = ivtmp_70 - _72;
> + if (ivtmp_74 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 13>; [16.67%]
> +
> + Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
> + rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
> + iteration amount for mutiple rgroups.
> +
> + The analysis of the flow of multiple rgroups:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + ...
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> +
> + We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
> + to be processed in each iteration.
> +
> + The related STOREs:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + Since these 2 STOREs store 2 vectors that the second vector is half elements
> + of the first vector. So the length of second STORE will be _64 = _72 / 2;
> + It's similar to the VIEW_CONVERT of handling masks in SLP.
> +
> + 3. Multiple rgroups for non-SLP auto-vectorization.
> +
> + # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
> + # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
> + # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
> + _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> + _29 = (void *) ivtmp.35_10;
> + _7 = &MEM <vector([4,4]) int> [(int *)_29];
> + vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
> + _33 = _29 + POLY_INT_CST [16, 16];
> + _34 = &MEM <vector([4,4]) int> [(int *)_33];
> + vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
> + vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
> + _30 = (void *) ivtmp.36_2;
> + _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
> + .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
> + ivtmp_27 = ivtmp_26 - _28;
> + ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
> + ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
> + if (ivtmp_27 != 0)
> + goto <bb 4>; [83.33%]
> + else
> + goto <bb 5>; [16.67%]
> +
> + The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> +
> + The length of first half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> +
> + The length of second half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> +
> + 1). _28 always <= POLY_INT_CST [8, 8].
> + 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
> + 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
> +*/
> +
> +static tree
> +vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
> + gimple_seq *preheader_seq,
> + gimple_seq *header_seq,
> + rgroup_controls *rgc, tree niters,
> + unsigned int controls_length)
> +{
> + tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + /* We are not allowing masked approach in SELECT_VL. */
> + gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
> +
> + tree ctrl_type = rgc->type;
> + unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
> + poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
> + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
> +
> + /* Calculate the maximum number of item values that the rgroup
> + handles in total, the number that it handles for each iteration
> + of the vector loop. */
> + tree nitems_total = niters;
> + if (nitems_per_iter != 1)
> + {
> + /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
> + these multiplications don't overflow. */
> + tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
> + nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
> + nitems_total, compare_factor);
> + }
> +
> + /* Convert the comparison value to the IV type (either a no-op or
> + a promotion). */
> + nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
> +
> + /* Create an induction variable that counts the number of items
> + processed. */
> + tree index_before_incr, index_after_incr;
> + gimple_stmt_iterator incr_gsi;
> + bool insert_after;
> + standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> +
> + /* Test the decremented IV, which will never underflow 0 since we have
> + IFN_SELECT_VL to gurantee that. */
> + tree test_limit = nitems_total;
> +
> + /* Provide a definition of each control in the group. */
> + tree ctrl;
> + unsigned int i;
> + FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
> + {
> + /* Previous controls will cover BIAS items. This control covers the
> + next batch. */
> + poly_uint64 bias = nitems_per_ctrl * i;
> + tree bias_tree = build_int_cst (iv_type, bias);
> +
> + /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
> + BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
> + control and adjust the bound down by BIAS. */
> + tree this_test_limit = test_limit;
> + if (i != 0)
> + {
> + this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + }
> +
> + /* Create decrement IV. */
> + create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
> + insert_after, &index_before_incr, &index_after_incr);
> +
> + poly_uint64 final_vf = vf * nitems_per_iter;
> + tree vf_step = build_int_cst (iv_type, final_vf);
> + tree res_len;
> + if (nitems_per_iter != 1 || controls_length != 1)
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
> +
> + For example:
> +
> + If non-final iteration process VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + ...
> +
> + If non-final iteration process non-VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + if (_71 % 2 == 0)
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + else
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
> + ...
> +
> + This is the simple case of 2-elements interleaved vector SLP. We
> + consider other interleave vector, the situation will become more
> + complicated. */
> + res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
> + index_before_incr, vf_step);
> + }
> + else
> + {
> + res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
> + index_before_incr, vf_step);
> + }
> + gassign *assign = gimple_build_assign (ctrl, res_len);
> + gimple_seq_add_stmt (header_seq, assign);
> + }
> +
> + return index_after_incr;
> +}
> +
> /* Set up the iteration condition and rgroup controls for LOOP, given
> that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
> loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
>
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
> unsigned int compare_precision = TYPE_PRECISION (compare_type);
> tree orig_niters = niters;
>
> @@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> continue;
> }
>
> + if (use_vl_p && rgc->controls.length () != 1)
> + {
> + rgroup_controls *sub_rgc
> + = &(*controls)[nmasks / rgc->controls.length () - 1];
> + if (!sub_rgc->controls.is_empty ()
> + && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
> + sub_rgc))
> + continue;
> + }
> +
> /* See whether zero-based IV would ever generate all-false masks
> or zero length before wrapping around. */
> bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
>
> /* Set up all controls for this group. */
> - test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
> - &preheader_seq,
> - &header_seq,
> - loop_cond_gsi, rgc,
> - niters, niters_skip,
> - might_wrap_p);
> + if (use_vl_p)
> + test_ctrl
> + = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
> + &preheader_seq, &header_seq,
> + rgc, niters, controls->length ());
> + else
> + test_ctrl
> + = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
> + &header_seq, loop_cond_gsi, rgc,
> + niters, niters_skip,
> + might_wrap_p);
Sorry for chiming in, I had some concern here.
We already have some handlings for partial vector in length in vect_set_loop_controls_directly
(actually it deals with both mask and length), the proposed vect_set_loop_controls_by_select_vl
for select_vl looks like a variant for partial vector in length (comparing to the normal MIN),
and also adopts decrement IV. IMHO, it seems better to modify/extend the current handling in
vect_set_loop_controls_directly for length, or factor out the handlings for length there and
extend the factored one. Otherwise, it means we have two sets of handlings for partial vector
in lengths, it looks bad to maintain. As the previous discussion, adopting decrement IV is an
enhancement for loop control, it's good for both cases w/ or w/o select_vl. If the above
understanding is correct, the possible steps seem to be:
- factor out the handling for length (* optional)
- modify it with decrement IV
- extend it with select_vl.
In future if some RVV vendor wants to degenerate select_vl to min, it can just adopt the same
handlings with min by not defining select_vl optab.
BR,
Kewen
> }
>
> /* Emit all accumulated statements. */
> @@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
> }
>
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (init, step, NULL_TREE, loop,
> + create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
> &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
> indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
> true, NULL_TREE, true,
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 6ea0f21fd13..d7cf92576d1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
> + create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
> insert_after, &indx_before_incr, &indx_after_incr);
>
> /* Next create a new phi node vector (NEW_PHI_TREE) which starts
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
>
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
>
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
>
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
>
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
>
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
>
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
>
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-08 6:27 ` juzhe.zhong
@ 2023-05-08 7:55 ` Kewen.Lin
2023-05-08 8:25 ` juzhe.zhong
0 siblings, 1 reply; 25+ messages in thread
From: Kewen.Lin @ 2023-05-08 7:55 UTC (permalink / raw)
To: juzhe.zhong; +Cc: gcc-patches, rguenther, richard.sandiford
Hi Juzhe,
> Hi, Kewen.
>
>>> Sorry for chiming in, I had some concern here.
>>> We already have some handlings for partial vector in length in vect_set_loop_controls_directly
>>>(actually it deals with both mask and length), the proposed vect_set_loop_controls_by_select_vl
>>>for select_vl looks like a variant for partial vector in length (comparing to the normal MIN),
>>>and also adopts decrement IV. IMHO, it seems better to modify/extend the current handling in
>>>vect_set_loop_controls_directly for length, or factor out the handlings for length there and
>>>extend the factored one. Otherwise, it means we have two sets of handlings for partial vector
>>>in lengths, it looks bad to maintain. As the previous discussion, adopting decrement IV is an
>>>enhancement for loop control, it's good for both cases w/ or w/o select_vl. If the above
>>>understanding is correct, the possible steps seem to be:
>>> - factor out the handling for length (* optional)
>>> - modify it with decrement IV
>>> - extend it with select_vl.
>>>In future if some RVV vendor wants to degenerate select_vl to min, it can just adopt the same
>>>handlings with min by not defining select_vl optab.
>
> You mean like this:
> doing this inside vect_set_loop_controls_directly ?
> if (use_while_len_p)
> return vect_set_loop_controls_by_while_len(...)
No, I meant either factoring out those handlings for partial vector in length in function
vect_set_loop_controls_directly to one separated function like: vect_set_loop_controls_directly_length
and rename the existing one to vect_set_loop_controls_directly_mask, or keep the existing
vect_set_loop_controls_directly for both mask and length but modify/extend the part for length.
If there is no much code to share between mask and length, the former may be better.
BR,
Kewen
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-08 7:55 ` Kewen.Lin
@ 2023-05-08 8:25 ` juzhe.zhong
0 siblings, 0 replies; 25+ messages in thread
From: juzhe.zhong @ 2023-05-08 8:25 UTC (permalink / raw)
To: linkw; +Cc: gcc-patches, rguenther, richard.sandiford
[-- Attachment #1: Type: text/plain, Size: 2822 bytes --]
Hi, Kewen.
I have tried to implement "decrement IV" feature and incorporate into "vect_set_loop_controls_directly".
Since the implementation is quite different from vect_set_loop_controls_directly, it will make vect_set_loop_controls_directly
very complicated sometimes it makes me very hard to debug when I am testing.
I am not sure but I can try again.
This patch isolated those implementation into a single function "vect_set_loop_controls_by_select_vl" which makes Richards
easier to review codes. Well, I think I can try again to incorporate those codes into "vect_set_loop_controls_directly" when they
finish the review process of "vect_set_loop_controls_by_select_vl".
Thanks.
juzhe.zhong@rivai.ai
From: Kewen.Lin
Date: 2023-05-08 15:55
To: juzhe.zhong@rivai.ai
CC: gcc-patches; rguenther; richard.sandiford
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
Hi Juzhe,
> Hi, Kewen.
>
>>> Sorry for chiming in, I had some concern here.
>>> We already have some handlings for partial vector in length in vect_set_loop_controls_directly
>>>(actually it deals with both mask and length), the proposed vect_set_loop_controls_by_select_vl
>>>for select_vl looks like a variant for partial vector in length (comparing to the normal MIN),
>>>and also adopts decrement IV. IMHO, it seems better to modify/extend the current handling in
>>>vect_set_loop_controls_directly for length, or factor out the handlings for length there and
>>>extend the factored one. Otherwise, it means we have two sets of handlings for partial vector
>>>in lengths, it looks bad to maintain. As the previous discussion, adopting decrement IV is an
>>>enhancement for loop control, it's good for both cases w/ or w/o select_vl. If the above
>>>understanding is correct, the possible steps seem to be:
>>> - factor out the handling for length (* optional)
>>> - modify it with decrement IV
>>> - extend it with select_vl.
>>>In future if some RVV vendor wants to degenerate select_vl to min, it can just adopt the same
>>>handlings with min by not defining select_vl optab.
>
> You mean like this:
> doing this inside vect_set_loop_controls_directly ?
> if (use_while_len_p)
> return vect_set_loop_controls_by_while_len(...)
No, I meant either factoring out those handlings for partial vector in length in function
vect_set_loop_controls_directly to one separated function like: vect_set_loop_controls_directly_length
and rename the existing one to vect_set_loop_controls_directly_mask, or keep the existing
vect_set_loop_controls_directly for both mask and length but modify/extend the part for length.
If there is no much code to share between mask and length, the former may be better.
BR,
Kewen
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-07 15:19 ` Jeff Law
2023-05-07 21:54 ` 钟居哲
@ 2023-05-09 10:52 ` juzhe.zhong
1 sibling, 0 replies; 25+ messages in thread
From: juzhe.zhong @ 2023-05-09 10:52 UTC (permalink / raw)
To: jeffreyalaw, gcc-patches; +Cc: richard.sandiford, rguenther
[-- Attachment #1: Type: text/plain, Size: 8160 bytes --]
Hi，Richards. Would you mind reviewing this patch?
Thanks.
juzhe.zhong@rivai.ai
From: Jeff Law
Date: 2023-05-07 23:19
To: juzhe.zhong; gcc-patches
CC: richard.sandiford; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
On 5/4/23 07:25, juzhe.zhong@rivai.ai wrote:
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
So at a high level this is pretty good. I think there's some
improvements we should make in the documentation and comments, but I'm
comfortable with most of the implementation details.
>
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
This reads rather poorly. Is this still accurate?
Set operand 0 to the number of active elements in a vector to be updated
in a loop iteration based on the total number of elements to be updated,
the vectorization factor and vector properties of the target.
> +operand 1 is the total elements need to be updated value.
operand 1 is the total elements in the vector to be updated.
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
This allows dynamic adjustment of the number of elements processed each
loop iteration. -- is that still accurate and does it read better?
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
It's been pretty standard to stick with just PLUS_EXPR for this stuff
and instead negate the constant to produce the same effect as
MINUS_EXPR. Is there a reason we're not continuing that practice?
Sorry if you've answered this already -- if you have, you can just point
me at the prior discussion and I'll read it.
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
I would suggest not using "permute" in this description. When I read
permute in the context of vectorization, I think of a vector permute to
scramble elements within a vector.
This looks like you're just adjusting how many vector elements you're
operating on.
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
Looks like you may have mixed up spaces and tabs in the above comment.
Just a nit, but let's go ahead and get it fixed.
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
>
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
When you break a line with a logical like this, go ahead and add
parenthesis and make sure the logical aligns just after the paren. ie
bool use_vl_p = (!use_masks_p
&& direct....
Alternately, compute the direct_itnernal_fn_supported_p into its own
boolean and then you don't need as much line wrapping.
In general, don't be afraid to use extra temporaries if doing so
improves readability.
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
This looks like a good example of code that would be easier to read if
the call to direct_internal-fn_supported_p was saved into a temporary.
Similarly for the instance you added in vectorizable_load.
I'd like to get this patch wrapped up soon. But I also want to give
both Richards a chance to chime in with their concerns.
Thanks,
Jeff
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-05 23:41 ` 钟居哲
@ 2023-05-09 12:59 ` Richard Sandiford
2023-05-09 13:27 ` juzhe.zhong
0 siblings, 1 reply; 25+ messages in thread
From: Richard Sandiford @ 2023-05-09 12:59 UTC (permalink / raw)
To: 钟居哲; +Cc: gcc-patches, rguenther
钟居哲 <juzhe.zhong@rivai.ai> writes:
> Hi, Richards. I would like to give more information about this patch so that it will make this patch easier for you to review.
>
> Currently, I saw we have 3 situations that we need to handle in case of loop control IV in auto-vectorization:
> 1. Single rgroup loop control (ncopies == 1 && vec_num == 1 so loop_len.length () == 1 or rgc->lengh () == 1)
> 2. Multiple rgroup for SLP.
> 3. Multiple rgroup for non-SLP which is Richard Sandiford point out previously (For example, VEC_PACK_TRUNC).
>
> To talk about this patch, let me talk about RVV LLVM implementation first which inspire me to send this patch:
> https://reviews.llvm.org/D99750
>
> According to LLVM implementation, they are adding a middle-end IR called "get_vector_length" which has totally
> same functionality as "select_vl" in this patch (I call it "while_len" previously, now I rename it as "select_vl" following Richard suggestion).
>
> The LLVM implementation is only let "get_vector_length" calculate the number of elements in single rgroup loop.
> For multi rgroup, let's take a look at it:
> https://godbolt.org/z/3GP78efTY
>
> void
> foo1 (short *__restrict f, int *__restrict d, int n)
> {
> for (int i = 0; i < n; ++i)
> {
> f[i * 2 + 0] = 1;
> f[i * 2 + 1] = 2;
> d[i] = 3;
> }
> }
>
> RISC-V Clang:
> foo1: # @foo1
> # %bb.0:
> blez a2, .LBB0_8
> # %bb.1:
> li a3, 16
> bgeu a2, a3, .LBB0_3
> # %bb.2:
> li a6, 0
> j .LBB0_6
> .LBB0_3:
> andi a6, a2, -16
> lui a3, 32
> addiw a3, a3, 1
> vsetivli zero, 8, e32, m2, ta, ma
> vmv.v.x v8, a3
> vmv.v.i v10, 3
> mv a4, a6
> mv a5, a1
> mv a3, a0
> .LBB0_4: # =>This Inner Loop Header: Depth=1
> addi a7, a5, 32
> addi t0, a3, 32
> vsetivli zero, 16, e16, m2, ta, ma
> vse16.v v8, (a3)
> vse16.v v8, (t0)
> vsetivli zero, 8, e32, m2, ta, ma
> vse32.v v10, (a5)
> vse32.v v10, (a7)
> addi a3, a3, 64
> addi a4, a4, -16
> addi a5, a5, 64
> bnez a4, .LBB0_4
> # %bb.5:
> beq a6, a2, .LBB0_8
> .LBB0_6:
> slli a3, a6, 2
> add a0, a0, a3
> addi a0, a0, 2
> add a1, a1, a3
> sub a2, a2, a6
> li a3, 1
> li a4, 2
> li a5, 3
> .LBB0_7: # =>This Inner Loop Header: Depth=1
> sh a3, -2(a0)
> sh a4, 0(a0)
> sw a5, 0(a1)
> addi a0, a0, 4
> addi a2, a2, -1
> addi a1, a1, 4
> bnez a2, .LBB0_7
> .LBB0_8:
> ret
>
> ARM GCC:
> foo1:
> cmp w2, 0
> ble .L1
> addvl x4, x0, #1
> mov x3, 0
> cntb x7
> cntb x6, all, mul #2
> sbfiz x2, x2, 1, 32
> ptrue p0.b, all
> mov x5, x2
> adrp x8, .LC0
> uqdech x5
> add x8, x8, :lo12:.LC0
> whilelo p1.h, xzr, x5
> ld1rw z1.s, p0/z, [x8]
> mov z0.s, #3
> whilelo p0.h, xzr, x2
> .L3:
> st1h z1.h, p0, [x0, x3, lsl 1]
> st1h z1.h, p1, [x4, x3, lsl 1]
> st1w z0.s, p1, [x1, #1, mul vl]
> add x3, x3, x7
> whilelo p1.h, x3, x5
> st1w z0.s, p0, [x1]
> add x1, x1, x6
> whilelo p0.h, x3, x2
> b.any .L3
> .L1:
> ret
>
> It's very obvious that ARM GCC has much better codegen since RVV LLVM just use SIMD style to handle multi-rgroup SLP auto-vectorization.
>
> Well, I am totally aggree that we should add length stuff in auto-vectorization not only for single rgroup but also multiple rgroup.
> However, when I am trying to implement multiple rgroup length for both SLP and non-SLP and testing, turns out it's hard to use select_vl
> since "select_vl" pattern allows non-VF flexible length (length <= min (remain,VF)) in any iteration, it's consuming much more operations for
> adjust loop controls IV and data reference address point IV than just using "MIN_EXPR".
>
> So for Case 2 && Case 3, I just use MIN_EXPR directly instead of SELECT_VL after my serveral internal testing.
Could you go into more details about this? I imagined that for case 3,
there would be a single SELECT_VL that decides how many scalar iterations
will be handled by the current vector iteration, then we would "expand"
the result (using MIN_EXPRs) to the multi-control cases.
In a sense that replicates what the SVE code above is doing. But for SVE,
it's possible to "optimise" the unpacking of a WHILELO result due to the
lack of implementation-defined behaviour. So conceptually we have a
single WHILELO that is unpacked one layer to give two masks. But in
reality we optimise that two WHILELOs. No such optimisation is possible
with SELECT_VL, and maybe that produces poor-quality code.
That might be what you mean (haven't had chance to look at the patch
itself yet, but hope to tomorrow).
For case 2 (max_nscalars_per_iter > 1), I think it would make conceptual
sense to pass max_nscalars_per_iter to SELECT_VL or (alternatively)
multiply the result of the SELECT_VL by max_nscalars_per_iter.
But it's only worth doing one of those two things if it produces
reasonable code for RVV.
> Now base on these situations, we only have "select_vl" for single-rgroup, but multiple-rgroup (both SLP and non-SLP), we just
> use MIN_EXPR.
>
> Is it more appropriate that we should remove "select_vl" and just use MIN_EXPR force VF elements in each non-final iteration in single rgroup?
It's up to you. If you don't think select_vl is worth it then it would
obviously make the vectoriser changes a bit simpler.
But making the vectoriser simpler isn't IMO the goal here. SELECT_VL
seems like a perfectly reasonable construct to add to target-independent
code. We just need to work out some of the details.
FWIW, I share Kewen's concern about duplicating too much logic between
masks, current lengths, and SELECT_VL lengths. But I haven't looked at
the patch yet and so I don't know how easy it would be to avoid that.
Thanks,
Richard
> Like the codegen according to RVV ISA example (show as RVV LLVM):
> https://repo.hca.bsc.es/epic/z/oynhzP
>
> ASM:
> vec_add: # @vec_add
> blez a3, .LBB0_3
> li a4, 0
> .LBB0_2: # %vector.body
> sub a5, a3, a4
> vsetvli a6, a5, e64, m1, ta, mu ==> change it into a6 = min (a5, VF) && vsetvli zero, a6, e64, m1, ta, mu
> slli a7, a4, 3
> add a5, a1, a7
> vle64.v v8, (a5)
> add a5, a2, a7
> vle64.v v9, (a5)
> vfadd.vv v8, v8, v9
> add a7, a7, a0
> add a4, a4, a6
> vse64.v v8, (a7)
> bne a4, a3, .LBB0_2
> .LBB0_3: # %for.cond.cleanup
> ret
>
> So if we remove "select_vl" just use MIN_EXPR to force VF elements in non-final iteration, we will end up with having like this:
>
> vec_add: # @vec_add
> blez a3, .LBB0_3
> csrr VF in bytes, vlenb
> .LBB0_2: # %vector.body
> sub a5, a3, VF in elements
> a6 = min (a5, VF in elements)
> vsetvli zero, a6, e64, m1, ta, mu
> add a5, a1, VF in bytes
> vle64.v v8, (a5)
> add a5, a2, VF in bytes
> vle64.v v9, (a5)
> vfadd.vv v8, v8, v9
> add a7, a7, a0
> vse64.v v8, (a7)
> bne a4, a3, .LBB0_2
> .LBB0_3: # %for.cond.cleanup
> ret
>
> Both codegen are working good for RVV.
> Just only the second one can not have the RVV special optimization (even distributed workload in last 2 iterations).
>
> Expecting any suggestions from you.
> Thanks.
>
>
> juzhe.zhong@rivai.ai
>
> From: juzhe.zhong
> Date: 2023-05-04 21:25
> To: gcc-patches
> CC: richard.sandiford; rguenther; Ju-Zhe Zhong
> Subject: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
> Thanks.
>
> ---
> gcc/cfgloopmanip.cc | 2 +-
> gcc/doc/md.texi | 34 +++
> gcc/gimple-loop-interchange.cc | 2 +-
> gcc/internal-fn.def | 1 +
> gcc/optabs.def | 1 +
> gcc/tree-ssa-loop-ivcanon.cc | 2 +-
> gcc/tree-ssa-loop-ivopts.cc | 2 +-
> gcc/tree-ssa-loop-manip.cc | 18 +-
> gcc/tree-ssa-loop-manip.h | 4 +-
> gcc/tree-vect-data-refs.cc | 8 +-
> gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
> gcc/tree-vect-loop.cc | 32 ++-
> gcc/tree-vect-stmts.cc | 89 +++++++-
> gcc/tree-vectorizer.h | 4 +-
> 14 files changed, 535 insertions(+), 38 deletions(-)
>
> diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
> index 0e3ad8ed742..6e09dcbb0b1 100644
> --- a/gcc/cfgloopmanip.cc
> +++ b/gcc/cfgloopmanip.cc
> @@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
> }
> gsi = gsi_last_bb (loop_header);
> - create_iv (initial_value, stride, iv, loop, &gsi, false,
> + create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
> iv_before, iv_after);
> /* Insert loop exit condition. */
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
> index 1b77bfd46b2..e5590374e59 100644
> --- a/gcc/gimple-loop-interchange.cc
> +++ b/gcc/gimple-loop-interchange.cc
> @@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
> tree var_before, var_after;
> tree base = unshare_expr (iv->init_expr);
> tree step = unshare_expr (iv->step);
> - create_iv (base, step, SSA_NAME_VAR (iv->var),
> + create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
> tgt.m_loop, &incr_pos, false, &var_before, &var_after);
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 7fe742c2ae7..6f6fa7d37f9 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
> DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
> DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
> +DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
> DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
> check_raw_ptrs, check_ptrs)
> DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 695f5911b30..b637471b76e 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
> OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
> OPTAB_D (len_load_optab, "len_load_$a")
> OPTAB_D (len_store_optab, "len_store_$a")
> +OPTAB_D (select_vl_optab, "select_vl$a")
> diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
> index e41ec73a52a..d4e113c44df 100644
> --- a/gcc/tree-ssa-loop-ivcanon.cc
> +++ b/gcc/tree-ssa-loop-ivcanon.cc
> @@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
> niter,
> build_int_cst (type, 1));
> incr_at = gsi_last_bb (in->src);
> - create_iv (niter,
> + create_iv (niter, PLUS_EXPR,
> build_int_cst (type, -1),
> NULL_TREE, loop,
> &incr_at, false, var_before, &var);
> diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
> index 78e8cbc75b5..8640fe2d487 100644
> --- a/gcc/tree-ssa-loop-ivopts.cc
> +++ b/gcc/tree-ssa-loop-ivopts.cc
> @@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
> base = unshare_expr (cand->iv->base);
> - create_iv (base, unshare_expr (cand->iv->step),
> + create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
> cand->var_before, data->current_loop,
> &incr_pos, after, &cand->var_before, &cand->var_after);
> }
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
> edge pe = loop_preheader_edge (loop);
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
> rewrite_all_phi_nodes_with_iv (loop, var_before);
> diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
> index d49273a3987..bda09f51d56 100644
> --- a/gcc/tree-ssa-loop-manip.h
> +++ b/gcc/tree-ssa-loop-manip.h
> @@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
> typedef void (*transform_callback)(class loop *, void *);
> -extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
> - bool, tree *, tree *);
> +extern void create_iv (tree, tree_code, tree, tree, class loop *,
> + gimple_stmt_iterator *, bool, tree *, tree *);
> extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
> extern void verify_loop_closed_ssa (bool, class loop * = NULL);
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index 6721ab6efc4..5c9103b16e5 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (aggr_ptr_init,
> + create_iv (aggr_ptr_init, PLUS_EXPR,
> fold_convert (aggr_ptr_type, iv_step),
> aggr_ptr, loop, &incr_gsi, insert_after,
> &indx_before_incr, &indx_after_incr);
> @@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> {
> standard_iv_increment_position (containing_loop, &incr_gsi,
> &insert_after);
> - create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
> - containing_loop, &incr_gsi, insert_after, &indx_before_incr,
> - &indx_after_incr);
> + create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
> + aggr_ptr, containing_loop, &incr_gsi, insert_after,
> + &indx_before_incr, &indx_after_incr);
> incr = gsi_stmt (incr_gsi);
> /* Copy the points-to information if it exists. */
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> @@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
> - &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
> + create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
> + loop, &incr_gsi, insert_after, &index_before_incr,
> + &index_after_incr);
> tree zero_index = build_int_cst (compare_type, 0);
> tree test_index, test_limit, first_limit;
> @@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> return next_ctrl;
> }
> +/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> + for all the rgroup controls in RGC and return a control that is nonzero
> + when the loop needs to iterate. Add any new preheader statements to
> + PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
> +
> + RGC belongs to loop LOOP. The loop originally iterated NITERS
> + times and has been vectorized according to LOOP_VINFO.
> +
> + Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
> + to TEST_LIMIT - bias.
> +
> + In vect_set_loop_controls_by_select_vl, we are iterating from start at
> + IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
> + IFN_SELECT_VL pattern.
> +
> + 1. Single rgroup, the Gimple IR should be:
> +
> + # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
> + # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
> + # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
> + # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
> + # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + ivtmp_15 = _28 * 4;
> + vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
> + _1 = B[i_10];
> + .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
> + i_7 = i_10 + 1;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
> + data reference pointer IV.
> +
> + 1). The result of .SELECT_VL:
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + The _28 is not necessary to be VF in any iteration, instead, we allow
> + _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
> + pattern allows target have various flexible optimizations in vector
> + loop iterations. Target like RISC-V has special application vector
> + length calculation instruction which will distribute even workload
> + in the last 2 iterations.
> +
> + Other example is that we can allow even generate _28 <= VF / 2 so
> + that some machine can run vector codes in low power mode.
> +
> + 2). Loop control IV:
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + This is the saturating-subtraction towards zero, the outcome of
> + .SELECT_VL wil make ivtmp_27 never underflow zero.
> +
> + 3). Data reference pointer IV:
> + ivtmp_15 = _28 * 4;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> +
> + The pointer IV is adjusted accurately according to the .SELECT_VL.
> +
> + 2. Multiple rgroup, the Gimple IR should be:
> +
> + # i_23 = PHI <i_20(6), 0(11)>
> + # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
> + # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
> + # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
> + # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + _1 = i_23 * 2;
> + _2 = (long unsigned int) _1;
> + _3 = _2 * 2;
> + _4 = f_15(D) + _3;
> + _5 = _2 + 1;
> + _6 = _5 * 2;
> + _7 = f_15(D) + _6;
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _8 = (long unsigned int) i_23;
> + _9 = _8 * 4;
> + _10 = d_18(D) + _9;
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + i_20 = i_23 + 1;
> + vectp_f.8_52 = vectp_f.8_56 + 16;
> + vectp_d.10_60 = vectp_d.10_63 + 16;
> + ivtmp_74 = ivtmp_73 - _75;
> + ivtmp_71 = ivtmp_70 - _72;
> + if (ivtmp_74 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 13>; [16.67%]
> +
> + Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
> + rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
> + iteration amount for mutiple rgroups.
> +
> + The analysis of the flow of multiple rgroups:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + ...
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> +
> + We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
> + to be processed in each iteration.
> +
> + The related STOREs:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + Since these 2 STOREs store 2 vectors that the second vector is half elements
> + of the first vector. So the length of second STORE will be _64 = _72 / 2;
> + It's similar to the VIEW_CONVERT of handling masks in SLP.
> +
> + 3. Multiple rgroups for non-SLP auto-vectorization.
> +
> + # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
> + # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
> + # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
> + _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> + _29 = (void *) ivtmp.35_10;
> + _7 = &MEM <vector([4,4]) int> [(int *)_29];
> + vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
> + _33 = _29 + POLY_INT_CST [16, 16];
> + _34 = &MEM <vector([4,4]) int> [(int *)_33];
> + vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
> + vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
> + _30 = (void *) ivtmp.36_2;
> + _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
> + .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
> + ivtmp_27 = ivtmp_26 - _28;
> + ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
> + ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
> + if (ivtmp_27 != 0)
> + goto <bb 4>; [83.33%]
> + else
> + goto <bb 5>; [16.67%]
> +
> + The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> +
> + The length of first half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> +
> + The length of second half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> +
> + 1). _28 always <= POLY_INT_CST [8, 8].
> + 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
> + 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
> +*/
> +
> +static tree
> +vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
> + gimple_seq *preheader_seq,
> + gimple_seq *header_seq,
> + rgroup_controls *rgc, tree niters,
> + unsigned int controls_length)
> +{
> + tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + /* We are not allowing masked approach in SELECT_VL. */
> + gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
> +
> + tree ctrl_type = rgc->type;
> + unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
> + poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
> + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
> +
> + /* Calculate the maximum number of item values that the rgroup
> + handles in total, the number that it handles for each iteration
> + of the vector loop. */
> + tree nitems_total = niters;
> + if (nitems_per_iter != 1)
> + {
> + /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
> + these multiplications don't overflow. */
> + tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
> + nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
> + nitems_total, compare_factor);
> + }
> +
> + /* Convert the comparison value to the IV type (either a no-op or
> + a promotion). */
> + nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
> +
> + /* Create an induction variable that counts the number of items
> + processed. */
> + tree index_before_incr, index_after_incr;
> + gimple_stmt_iterator incr_gsi;
> + bool insert_after;
> + standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> +
> + /* Test the decremented IV, which will never underflow 0 since we have
> + IFN_SELECT_VL to gurantee that. */
> + tree test_limit = nitems_total;
> +
> + /* Provide a definition of each control in the group. */
> + tree ctrl;
> + unsigned int i;
> + FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
> + {
> + /* Previous controls will cover BIAS items. This control covers the
> + next batch. */
> + poly_uint64 bias = nitems_per_ctrl * i;
> + tree bias_tree = build_int_cst (iv_type, bias);
> +
> + /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
> + BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
> + control and adjust the bound down by BIAS. */
> + tree this_test_limit = test_limit;
> + if (i != 0)
> + {
> + this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + }
> +
> + /* Create decrement IV. */
> + create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
> + insert_after, &index_before_incr, &index_after_incr);
> +
> + poly_uint64 final_vf = vf * nitems_per_iter;
> + tree vf_step = build_int_cst (iv_type, final_vf);
> + tree res_len;
> + if (nitems_per_iter != 1 || controls_length != 1)
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
> +
> + For example:
> +
> + If non-final iteration process VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + ...
> +
> + If non-final iteration process non-VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + if (_71 % 2 == 0)
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + else
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
> + ...
> +
> + This is the simple case of 2-elements interleaved vector SLP. We
> + consider other interleave vector, the situation will become more
> + complicated. */
> + res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
> + index_before_incr, vf_step);
> + }
> + else
> + {
> + res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
> + index_before_incr, vf_step);
> + }
> + gassign *assign = gimple_build_assign (ctrl, res_len);
> + gimple_seq_add_stmt (header_seq, assign);
> + }
> +
> + return index_after_incr;
> +}
> +
> /* Set up the iteration condition and rgroup controls for LOOP, given
> that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
> loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
> unsigned int compare_precision = TYPE_PRECISION (compare_type);
> tree orig_niters = niters;
> @@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> continue;
> }
> + if (use_vl_p && rgc->controls.length () != 1)
> + {
> + rgroup_controls *sub_rgc
> + = &(*controls)[nmasks / rgc->controls.length () - 1];
> + if (!sub_rgc->controls.is_empty ()
> + && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
> + sub_rgc))
> + continue;
> + }
> +
> /* See whether zero-based IV would ever generate all-false masks
> or zero length before wrapping around. */
> bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
> /* Set up all controls for this group. */
> - test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
> - &preheader_seq,
> - &header_seq,
> - loop_cond_gsi, rgc,
> - niters, niters_skip,
> - might_wrap_p);
> + if (use_vl_p)
> + test_ctrl
> + = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
> + &preheader_seq, &header_seq,
> + rgc, niters, controls->length ());
> + else
> + test_ctrl
> + = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
> + &header_seq, loop_cond_gsi, rgc,
> + niters, niters_skip,
> + might_wrap_p);
> }
> /* Emit all accumulated statements. */
> @@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
> }
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (init, step, NULL_TREE, loop,
> + create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
> &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
> indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
> true, NULL_TREE, true,
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 6ea0f21fd13..d7cf92576d1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
> + create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
> insert_after, &indx_before_incr, &indx_after_incr);
> /* Next create a new phi node vector (NEW_PHI_TREE) which starts
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-09 12:59 ` Richard Sandiford
@ 2023-05-09 13:27 ` juzhe.zhong
2023-05-09 14:34 ` 钟居哲
0 siblings, 1 reply; 25+ messages in thread
From: juzhe.zhong @ 2023-05-09 13:27 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 55160 bytes --]
Thanks, Richard.
>> Could you go into more details about this? I imagined that for case 3,
>> there would be a single SELECT_VL that decides how many scalar iterations
>> will be handled by the current vector iteration, then we would "expand"
>> the result (using MIN_EXPRs) to the multi-control cases.
For case 2 , here is the example:
+ 2. Multiple rgroup, the Gimple IR should be:
+
+ # i_23 = PHI <i_20(6), 0(11)>
+ # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
+ # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
+ # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
+ # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ _1 = i_23 * 2;
+ _2 = (long unsigned int) _1;
+ _3 = _2 * 2;
+ _4 = f_15(D) + _3;
+ _5 = _2 + 1;
+ _6 = _5 * 2;
+ _7 = f_15(D) + _6;
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ _8 = (long unsigned int) i_23;
+ _9 = _8 * 4;
+ _10 = d_18(D) + _9;
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+ i_20 = i_23 + 1;
+ vectp_f.8_52 = vectp_f.8_56 + 16;
+ vectp_d.10_60 = vectp_d.10_63 + 16;
+ ivtmp_74 = ivtmp_73 - _75;
+ ivtmp_71 = ivtmp_70 - _72;
+ if (ivtmp_74 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 13>; [16.67%]
+
+ Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
+ rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
+ iteration amount for mutiple rgroups.+ The analysis of the flow of multiple rgroups:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ ...
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);Here, If use SELECT_VL instead of MIN_EXPR. Since we define the outcome of SELECT_VL can be any number in non-final iteration.It seems not easy to adjust address pointer IV (vectp_f.8_56 = vectp_f.8_51 + 16;) and the next length (_61 = _75 / 2;).
For case 3: + 3. Multiple rgroups for non-SLP auto-vectorization.+
+ # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
+ # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
+ # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
+ _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+ _29 = (void *) ivtmp.35_10;
+ _7 = &MEM <vector([4,4]) int> [(int *)_29];
+ vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
+ _33 = _29 + POLY_INT_CST [16, 16];
+ _34 = &MEM <vector([4,4]) int> [(int *)_33];
+ vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
+ vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
+ _30 = (void *) ivtmp.36_2;
+ _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
+ .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
+ ivtmp_27 = ivtmp_26 - _28;
+ ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
+ ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
+ if (ivtmp_27 != 0)
+ goto <bb 4>; [83.33%]
+ else
+ goto <bb 5>; [16.67%]
+
+ The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+
+ The length of first half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+
+ The length of second half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+
+ 1). _28 always <= POLY_INT_CST [8, 8].
+ 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
+ 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.We known in Case 3, we should deal with 2 vectors: vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;First we use _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>; to generate the number of elements to be processedfor these 2 vector.Second, we use "loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;" "loop_len_15" is the number elements to be processed for first vector.Then, "loop_len_16 = _28 - loop_len_15; "loop_len_16" is the number elements to be processed for first vector.I think "loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;" is very similiar the unpacklo in ARM SVE."loop_len_16 = _28 - loop_len_15; "loop_len_16" is very similiar the unpackhi in ARM SVE.
>> It's up to you. If you don't think select_vl is worth it then it would
>>obviously make the vectoriser changes a bit simpler.
>>But making the vectoriser simpler isn't IMO the goal here. SELECT_VL
>>seems like a perfectly reasonable construct to add to target-independent
>>code. We just need to work out some of the details.
Ok, I also prefer keeping select_vl.
>>FWIW, I share Kewen's concern about duplicating too much logic between
>>masks, current lengths, and SELECT_VL lengths. But I haven't looked at
>>the patch yet and so I don't know how easy it would be to avoid that.
I understand the concern, the current implementation are in the isolated function "vect_set_loop_controls_by_select_vl",
it's easier to review the implementation.
Maybe we can first make the whole implementation codes in "vect_set_loop_controls_by_select_vl" to be stable after review,
then we can try to incorporate these codes of "vect_set_loop_controls_by_select_vl" into "vect_set_loop_controls_directly".
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-09 20:59
To: 钟居哲
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
钟居哲 <juzhe.zhong@rivai.ai> writes:
> Hi, Richards. I would like to give more information about this patch so that it will make this patch easier for you to review.
>
> Currently, I saw we have 3 situations that we need to handle in case of loop control IV in auto-vectorization:
> 1. Single rgroup loop control (ncopies == 1 && vec_num == 1 so loop_len.length () == 1 or rgc->lengh () == 1)
> 2. Multiple rgroup for SLP.
> 3. Multiple rgroup for non-SLP which is Richard Sandiford point out previously (For example, VEC_PACK_TRUNC).
>
> To talk about this patch, let me talk about RVV LLVM implementation first which inspire me to send this patch:
> https://reviews.llvm.org/D99750
>
> According to LLVM implementation, they are adding a middle-end IR called "get_vector_length" which has totally
> same functionality as "select_vl" in this patch (I call it "while_len" previously, now I rename it as "select_vl" following Richard suggestion).
>
> The LLVM implementation is only let "get_vector_length" calculate the number of elements in single rgroup loop.
> For multi rgroup, let's take a look at it:
> https://godbolt.org/z/3GP78efTY
>
> void
> foo1 (short *__restrict f, int *__restrict d, int n)
> {
> for (int i = 0; i < n; ++i)
> {
> f[i * 2 + 0] = 1;
> f[i * 2 + 1] = 2;
> d[i] = 3;
> }
> }
>
> RISC-V Clang:
> foo1: # @foo1
> # %bb.0:
> blez a2, .LBB0_8
> # %bb.1:
> li a3, 16
> bgeu a2, a3, .LBB0_3
> # %bb.2:
> li a6, 0
> j .LBB0_6
> .LBB0_3:
> andi a6, a2, -16
> lui a3, 32
> addiw a3, a3, 1
> vsetivli zero, 8, e32, m2, ta, ma
> vmv.v.x v8, a3
> vmv.v.i v10, 3
> mv a4, a6
> mv a5, a1
> mv a3, a0
> .LBB0_4: # =>This Inner Loop Header: Depth=1
> addi a7, a5, 32
> addi t0, a3, 32
> vsetivli zero, 16, e16, m2, ta, ma
> vse16.v v8, (a3)
> vse16.v v8, (t0)
> vsetivli zero, 8, e32, m2, ta, ma
> vse32.v v10, (a5)
> vse32.v v10, (a7)
> addi a3, a3, 64
> addi a4, a4, -16
> addi a5, a5, 64
> bnez a4, .LBB0_4
> # %bb.5:
> beq a6, a2, .LBB0_8
> .LBB0_6:
> slli a3, a6, 2
> add a0, a0, a3
> addi a0, a0, 2
> add a1, a1, a3
> sub a2, a2, a6
> li a3, 1
> li a4, 2
> li a5, 3
> .LBB0_7: # =>This Inner Loop Header: Depth=1
> sh a3, -2(a0)
> sh a4, 0(a0)
> sw a5, 0(a1)
> addi a0, a0, 4
> addi a2, a2, -1
> addi a1, a1, 4
> bnez a2, .LBB0_7
> .LBB0_8:
> ret
>
> ARM GCC:
> foo1:
> cmp w2, 0
> ble .L1
> addvl x4, x0, #1
> mov x3, 0
> cntb x7
> cntb x6, all, mul #2
> sbfiz x2, x2, 1, 32
> ptrue p0.b, all
> mov x5, x2
> adrp x8, .LC0
> uqdech x5
> add x8, x8, :lo12:.LC0
> whilelo p1.h, xzr, x5
> ld1rw z1.s, p0/z, [x8]
> mov z0.s, #3
> whilelo p0.h, xzr, x2
> .L3:
> st1h z1.h, p0, [x0, x3, lsl 1]
> st1h z1.h, p1, [x4, x3, lsl 1]
> st1w z0.s, p1, [x1, #1, mul vl]
> add x3, x3, x7
> whilelo p1.h, x3, x5
> st1w z0.s, p0, [x1]
> add x1, x1, x6
> whilelo p0.h, x3, x2
> b.any .L3
> .L1:
> ret
>
> It's very obvious that ARM GCC has much better codegen since RVV LLVM just use SIMD style to handle multi-rgroup SLP auto-vectorization.
>
> Well, I am totally aggree that we should add length stuff in auto-vectorization not only for single rgroup but also multiple rgroup.
> However, when I am trying to implement multiple rgroup length for both SLP and non-SLP and testing, turns out it's hard to use select_vl
> since "select_vl" pattern allows non-VF flexible length (length <= min (remain,VF)) in any iteration, it's consuming much more operations for
> adjust loop controls IV and data reference address point IV than just using "MIN_EXPR".
>
> So for Case 2 && Case 3, I just use MIN_EXPR directly instead of SELECT_VL after my serveral internal testing.
Could you go into more details about this? I imagined that for case 3,
there would be a single SELECT_VL that decides how many scalar iterations
will be handled by the current vector iteration, then we would "expand"
the result (using MIN_EXPRs) to the multi-control cases.
In a sense that replicates what the SVE code above is doing. But for SVE,
it's possible to "optimise" the unpacking of a WHILELO result due to the
lack of implementation-defined behaviour. So conceptually we have a
single WHILELO that is unpacked one layer to give two masks. But in
reality we optimise that two WHILELOs. No such optimisation is possible
with SELECT_VL, and maybe that produces poor-quality code.
That might be what you mean (haven't had chance to look at the patch
itself yet, but hope to tomorrow).
For case 2 (max_nscalars_per_iter > 1), I think it would make conceptual
sense to pass max_nscalars_per_iter to SELECT_VL or (alternatively)
multiply the result of the SELECT_VL by max_nscalars_per_iter.
But it's only worth doing one of those two things if it produces
reasonable code for RVV.
> Now base on these situations, we only have "select_vl" for single-rgroup, but multiple-rgroup (both SLP and non-SLP), we just
> use MIN_EXPR.
>
> Is it more appropriate that we should remove "select_vl" and just use MIN_EXPR force VF elements in each non-final iteration in single rgroup?
It's up to you. If you don't think select_vl is worth it then it would
obviously make the vectoriser changes a bit simpler.
But making the vectoriser simpler isn't IMO the goal here. SELECT_VL
seems like a perfectly reasonable construct to add to target-independent
code. We just need to work out some of the details.
FWIW, I share Kewen's concern about duplicating too much logic between
masks, current lengths, and SELECT_VL lengths. But I haven't looked at
the patch yet and so I don't know how easy it would be to avoid that.
Thanks,
Richard
> Like the codegen according to RVV ISA example (show as RVV LLVM):
> https://repo.hca.bsc.es/epic/z/oynhzP
>
> ASM:
> vec_add: # @vec_add
> blez a3, .LBB0_3
> li a4, 0
> .LBB0_2: # %vector.body
> sub a5, a3, a4
> vsetvli a6, a5, e64, m1, ta, mu ==> change it into a6 = min (a5, VF) && vsetvli zero, a6, e64, m1, ta, mu
> slli a7, a4, 3
> add a5, a1, a7
> vle64.v v8, (a5)
> add a5, a2, a7
> vle64.v v9, (a5)
> vfadd.vv v8, v8, v9
> add a7, a7, a0
> add a4, a4, a6
> vse64.v v8, (a7)
> bne a4, a3, .LBB0_2
> .LBB0_3: # %for.cond.cleanup
> ret
>
> So if we remove "select_vl" just use MIN_EXPR to force VF elements in non-final iteration, we will end up with having like this:
>
> vec_add: # @vec_add
> blez a3, .LBB0_3
> csrr VF in bytes, vlenb
> .LBB0_2: # %vector.body
> sub a5, a3, VF in elements
> a6 = min (a5, VF in elements)
> vsetvli zero, a6, e64, m1, ta, mu
> add a5, a1, VF in bytes
> vle64.v v8, (a5)
> add a5, a2, VF in bytes
> vle64.v v9, (a5)
> vfadd.vv v8, v8, v9
> add a7, a7, a0
> vse64.v v8, (a7)
> bne a4, a3, .LBB0_2
> .LBB0_3: # %for.cond.cleanup
> ret
>
> Both codegen are working good for RVV.
> Just only the second one can not have the RVV special optimization (even distributed workload in last 2 iterations).
>
> Expecting any suggestions from you.
> Thanks.
>
>
> juzhe.zhong@rivai.ai
>
> From: juzhe.zhong
> Date: 2023-05-04 21:25
> To: gcc-patches
> CC: richard.sandiford; rguenther; Ju-Zhe Zhong
> Subject: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
> Thanks.
>
> ---
> gcc/cfgloopmanip.cc | 2 +-
> gcc/doc/md.texi | 34 +++
> gcc/gimple-loop-interchange.cc | 2 +-
> gcc/internal-fn.def | 1 +
> gcc/optabs.def | 1 +
> gcc/tree-ssa-loop-ivcanon.cc | 2 +-
> gcc/tree-ssa-loop-ivopts.cc | 2 +-
> gcc/tree-ssa-loop-manip.cc | 18 +-
> gcc/tree-ssa-loop-manip.h | 4 +-
> gcc/tree-vect-data-refs.cc | 8 +-
> gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
> gcc/tree-vect-loop.cc | 32 ++-
> gcc/tree-vect-stmts.cc | 89 +++++++-
> gcc/tree-vectorizer.h | 4 +-
> 14 files changed, 535 insertions(+), 38 deletions(-)
>
> diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
> index 0e3ad8ed742..6e09dcbb0b1 100644
> --- a/gcc/cfgloopmanip.cc
> +++ b/gcc/cfgloopmanip.cc
> @@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
> }
> gsi = gsi_last_bb (loop_header);
> - create_iv (initial_value, stride, iv, loop, &gsi, false,
> + create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
> iv_before, iv_after);
> /* Insert loop exit condition. */
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
> index 1b77bfd46b2..e5590374e59 100644
> --- a/gcc/gimple-loop-interchange.cc
> +++ b/gcc/gimple-loop-interchange.cc
> @@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
> tree var_before, var_after;
> tree base = unshare_expr (iv->init_expr);
> tree step = unshare_expr (iv->step);
> - create_iv (base, step, SSA_NAME_VAR (iv->var),
> + create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
> tgt.m_loop, &incr_pos, false, &var_before, &var_after);
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 7fe742c2ae7..6f6fa7d37f9 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
> DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
> DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
> +DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
> DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
> check_raw_ptrs, check_ptrs)
> DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 695f5911b30..b637471b76e 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
> OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
> OPTAB_D (len_load_optab, "len_load_$a")
> OPTAB_D (len_store_optab, "len_store_$a")
> +OPTAB_D (select_vl_optab, "select_vl$a")
> diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
> index e41ec73a52a..d4e113c44df 100644
> --- a/gcc/tree-ssa-loop-ivcanon.cc
> +++ b/gcc/tree-ssa-loop-ivcanon.cc
> @@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
> niter,
> build_int_cst (type, 1));
> incr_at = gsi_last_bb (in->src);
> - create_iv (niter,
> + create_iv (niter, PLUS_EXPR,
> build_int_cst (type, -1),
> NULL_TREE, loop,
> &incr_at, false, var_before, &var);
> diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
> index 78e8cbc75b5..8640fe2d487 100644
> --- a/gcc/tree-ssa-loop-ivopts.cc
> +++ b/gcc/tree-ssa-loop-ivopts.cc
> @@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
> base = unshare_expr (cand->iv->base);
> - create_iv (base, unshare_expr (cand->iv->step),
> + create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
> cand->var_before, data->current_loop,
> &incr_pos, after, &cand->var_before, &cand->var_after);
> }
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
> edge pe = loop_preheader_edge (loop);
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
> rewrite_all_phi_nodes_with_iv (loop, var_before);
> diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
> index d49273a3987..bda09f51d56 100644
> --- a/gcc/tree-ssa-loop-manip.h
> +++ b/gcc/tree-ssa-loop-manip.h
> @@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
> typedef void (*transform_callback)(class loop *, void *);
> -extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
> - bool, tree *, tree *);
> +extern void create_iv (tree, tree_code, tree, tree, class loop *,
> + gimple_stmt_iterator *, bool, tree *, tree *);
> extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
> extern void verify_loop_closed_ssa (bool, class loop * = NULL);
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index 6721ab6efc4..5c9103b16e5 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (aggr_ptr_init,
> + create_iv (aggr_ptr_init, PLUS_EXPR,
> fold_convert (aggr_ptr_type, iv_step),
> aggr_ptr, loop, &incr_gsi, insert_after,
> &indx_before_incr, &indx_after_incr);
> @@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> {
> standard_iv_increment_position (containing_loop, &incr_gsi,
> &insert_after);
> - create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
> - containing_loop, &incr_gsi, insert_after, &indx_before_incr,
> - &indx_after_incr);
> + create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
> + aggr_ptr, containing_loop, &incr_gsi, insert_after,
> + &indx_before_incr, &indx_after_incr);
> incr = gsi_stmt (incr_gsi);
> /* Copy the points-to information if it exists. */
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> @@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
> - &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
> + create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
> + loop, &incr_gsi, insert_after, &index_before_incr,
> + &index_after_incr);
> tree zero_index = build_int_cst (compare_type, 0);
> tree test_index, test_limit, first_limit;
> @@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> return next_ctrl;
> }
> +/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> + for all the rgroup controls in RGC and return a control that is nonzero
> + when the loop needs to iterate. Add any new preheader statements to
> + PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
> +
> + RGC belongs to loop LOOP. The loop originally iterated NITERS
> + times and has been vectorized according to LOOP_VINFO.
> +
> + Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
> + to TEST_LIMIT - bias.
> +
> + In vect_set_loop_controls_by_select_vl, we are iterating from start at
> + IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
> + IFN_SELECT_VL pattern.
> +
> + 1. Single rgroup, the Gimple IR should be:
> +
> + # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
> + # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
> + # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
> + # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
> + # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + ivtmp_15 = _28 * 4;
> + vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
> + _1 = B[i_10];
> + .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
> + i_7 = i_10 + 1;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
> + data reference pointer IV.
> +
> + 1). The result of .SELECT_VL:
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + The _28 is not necessary to be VF in any iteration, instead, we allow
> + _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
> + pattern allows target have various flexible optimizations in vector
> + loop iterations. Target like RISC-V has special application vector
> + length calculation instruction which will distribute even workload
> + in the last 2 iterations.
> +
> + Other example is that we can allow even generate _28 <= VF / 2 so
> + that some machine can run vector codes in low power mode.
> +
> + 2). Loop control IV:
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + This is the saturating-subtraction towards zero, the outcome of
> + .SELECT_VL wil make ivtmp_27 never underflow zero.
> +
> + 3). Data reference pointer IV:
> + ivtmp_15 = _28 * 4;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> +
> + The pointer IV is adjusted accurately according to the .SELECT_VL.
> +
> + 2. Multiple rgroup, the Gimple IR should be:
> +
> + # i_23 = PHI <i_20(6), 0(11)>
> + # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
> + # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
> + # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
> + # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + _1 = i_23 * 2;
> + _2 = (long unsigned int) _1;
> + _3 = _2 * 2;
> + _4 = f_15(D) + _3;
> + _5 = _2 + 1;
> + _6 = _5 * 2;
> + _7 = f_15(D) + _6;
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _8 = (long unsigned int) i_23;
> + _9 = _8 * 4;
> + _10 = d_18(D) + _9;
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + i_20 = i_23 + 1;
> + vectp_f.8_52 = vectp_f.8_56 + 16;
> + vectp_d.10_60 = vectp_d.10_63 + 16;
> + ivtmp_74 = ivtmp_73 - _75;
> + ivtmp_71 = ivtmp_70 - _72;
> + if (ivtmp_74 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 13>; [16.67%]
> +
> + Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
> + rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
> + iteration amount for mutiple rgroups.
> +
> + The analysis of the flow of multiple rgroups:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + ...
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> +
> + We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
> + to be processed in each iteration.
> +
> + The related STOREs:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + Since these 2 STOREs store 2 vectors that the second vector is half elements
> + of the first vector. So the length of second STORE will be _64 = _72 / 2;
> + It's similar to the VIEW_CONVERT of handling masks in SLP.
> +
> + 3. Multiple rgroups for non-SLP auto-vectorization.
> +
> + # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
> + # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
> + # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
> + _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> + _29 = (void *) ivtmp.35_10;
> + _7 = &MEM <vector([4,4]) int> [(int *)_29];
> + vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
> + _33 = _29 + POLY_INT_CST [16, 16];
> + _34 = &MEM <vector([4,4]) int> [(int *)_33];
> + vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
> + vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
> + _30 = (void *) ivtmp.36_2;
> + _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
> + .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
> + ivtmp_27 = ivtmp_26 - _28;
> + ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
> + ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
> + if (ivtmp_27 != 0)
> + goto <bb 4>; [83.33%]
> + else
> + goto <bb 5>; [16.67%]
> +
> + The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> +
> + The length of first half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> +
> + The length of second half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> +
> + 1). _28 always <= POLY_INT_CST [8, 8].
> + 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
> + 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
> +*/
> +
> +static tree
> +vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
> + gimple_seq *preheader_seq,
> + gimple_seq *header_seq,
> + rgroup_controls *rgc, tree niters,
> + unsigned int controls_length)
> +{
> + tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + /* We are not allowing masked approach in SELECT_VL. */
> + gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
> +
> + tree ctrl_type = rgc->type;
> + unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
> + poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
> + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
> +
> + /* Calculate the maximum number of item values that the rgroup
> + handles in total, the number that it handles for each iteration
> + of the vector loop. */
> + tree nitems_total = niters;
> + if (nitems_per_iter != 1)
> + {
> + /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
> + these multiplications don't overflow. */
> + tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
> + nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
> + nitems_total, compare_factor);
> + }
> +
> + /* Convert the comparison value to the IV type (either a no-op or
> + a promotion). */
> + nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
> +
> + /* Create an induction variable that counts the number of items
> + processed. */
> + tree index_before_incr, index_after_incr;
> + gimple_stmt_iterator incr_gsi;
> + bool insert_after;
> + standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> +
> + /* Test the decremented IV, which will never underflow 0 since we have
> + IFN_SELECT_VL to gurantee that. */
> + tree test_limit = nitems_total;
> +
> + /* Provide a definition of each control in the group. */
> + tree ctrl;
> + unsigned int i;
> + FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
> + {
> + /* Previous controls will cover BIAS items. This control covers the
> + next batch. */
> + poly_uint64 bias = nitems_per_ctrl * i;
> + tree bias_tree = build_int_cst (iv_type, bias);
> +
> + /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
> + BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
> + control and adjust the bound down by BIAS. */
> + tree this_test_limit = test_limit;
> + if (i != 0)
> + {
> + this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + }
> +
> + /* Create decrement IV. */
> + create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
> + insert_after, &index_before_incr, &index_after_incr);
> +
> + poly_uint64 final_vf = vf * nitems_per_iter;
> + tree vf_step = build_int_cst (iv_type, final_vf);
> + tree res_len;
> + if (nitems_per_iter != 1 || controls_length != 1)
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
> +
> + For example:
> +
> + If non-final iteration process VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + ...
> +
> + If non-final iteration process non-VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + if (_71 % 2 == 0)
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + else
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
> + ...
> +
> + This is the simple case of 2-elements interleaved vector SLP. We
> + consider other interleave vector, the situation will become more
> + complicated. */
> + res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
> + index_before_incr, vf_step);
> + }
> + else
> + {
> + res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
> + index_before_incr, vf_step);
> + }
> + gassign *assign = gimple_build_assign (ctrl, res_len);
> + gimple_seq_add_stmt (header_seq, assign);
> + }
> +
> + return index_after_incr;
> +}
> +
> /* Set up the iteration condition and rgroup controls for LOOP, given
> that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
> loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
> unsigned int compare_precision = TYPE_PRECISION (compare_type);
> tree orig_niters = niters;
> @@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> continue;
> }
> + if (use_vl_p && rgc->controls.length () != 1)
> + {
> + rgroup_controls *sub_rgc
> + = &(*controls)[nmasks / rgc->controls.length () - 1];
> + if (!sub_rgc->controls.is_empty ()
> + && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
> + sub_rgc))
> + continue;
> + }
> +
> /* See whether zero-based IV would ever generate all-false masks
> or zero length before wrapping around. */
> bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
> /* Set up all controls for this group. */
> - test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
> - &preheader_seq,
> - &header_seq,
> - loop_cond_gsi, rgc,
> - niters, niters_skip,
> - might_wrap_p);
> + if (use_vl_p)
> + test_ctrl
> + = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
> + &preheader_seq, &header_seq,
> + rgc, niters, controls->length ());
> + else
> + test_ctrl
> + = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
> + &header_seq, loop_cond_gsi, rgc,
> + niters, niters_skip,
> + might_wrap_p);
> }
> /* Emit all accumulated statements. */
> @@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
> }
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (init, step, NULL_TREE, loop,
> + create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
> &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
> indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
> true, NULL_TREE, true,
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 6ea0f21fd13..d7cf92576d1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
> + create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
> insert_after, &indx_before_incr, &indx_after_incr);
> /* Next create a new phi node vector (NEW_PHI_TREE) which starts
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-09 13:27 ` juzhe.zhong
@ 2023-05-09 14:34 ` 钟居哲
0 siblings, 0 replies; 25+ messages in thread
From: 钟居哲 @ 2023-05-09 14:34 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 59240 bytes --]
More details for Case 2:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ ...
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
You may be confused that _61 = _75 / 2; and _64 = _72 / 2;
Well, this is similiar VIEW_CONVERT_EXPR of mask in ARM SVE.
Like ARM SVE:
tree
vect_get_loop_mask (gimple_stmt_iterator *gsi, vec_loop_masks *masks,
unsigned int nvectors, tree vectype, unsigned int index)
{
rgroup_controls *rgm = &(*masks)[nvectors - 1];
tree mask_type = rgm->type;
/* Populate the rgroup's mask array, if this is the first time we've
used it. */
if (rgm->controls.is_empty ())
{
rgm->controls.safe_grow_cleared (nvectors, true);
for (unsigned int i = 0; i < nvectors; ++i)
{
tree mask = make_temp_ssa_name (mask_type, NULL, "loop_mask");
/* Provide a dummy definition until the real one is available. */
SSA_NAME_DEF_STMT (mask) = gimple_build_nop ();
rgm->controls[i] = mask;
}
}
tree mask = rgm->controls[index];
if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type),
TYPE_VECTOR_SUBPARTS (vectype)))
{
/* A loop mask for data type X can be reused for data type Y
if X has N times more elements than Y and if Y's elements
are N times bigger than X's. In this case each sequence
of N elements in the loop mask will be all-zero or all-one.
We can then view-convert the mask so that each sequence of
N elements is replaced by a single element. */
gcc_assert (multiple_p (TYPE_VECTOR_SUBPARTS (mask_type),
TYPE_VECTOR_SUBPARTS (vectype)));
gimple_seq seq = NULL;
mask_type = truth_type_for (vectype);
mask = gimple_build (&seq, VIEW_CONVERT_EXPR, mask_type, mask);
if (seq)
gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
}
return mask;
}
I am doing similiar thing for RVV:
+vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
+ vec_loop_lens *lens, unsigned int nvectors, tree vectype,
+ unsigned int index)
{
rgroup_controls *rgl = &(*lens)[nvectors - 1];
bool use_bias_adjusted_len =
LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
+ tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
/* Populate the rgroup's len array, if this is the first time we've
used it. */
@@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
if (use_bias_adjusted_len)
return rgl->bias_adjusted_ctrl;
+ else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
+ OPTIMIZE_FOR_SPEED))
+ {
+ tree loop_len = rgl->controls[index];
+ poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
+ poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
+ if (maybe_ne (nunits1, nunits2))
+ {
+ /* A loop len for data type X can be reused for data type Y
+ if X has N times more elements than Y and if Y's elements
+ are N times bigger than X's. */
+ gcc_assert (multiple_p (nunits1, nunits2));
+ unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
+ gimple_seq seq = NULL;
+ loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
+ build_int_cst (iv_type, factor));
+ if (seq)
+ gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
+ }
+ return loop_len;
+ }
else
return rgl->controls[index];
}Thansk.
juzhe.zhong@rivai.ai
From: juzhe.zhong@rivai.ai
Date: 2023-05-09 21:27
To: richard.sandiford
CC: gcc-patches; rguenther
Subject: Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
Thanks, Richard.
>> Could you go into more details about this? I imagined that for case 3,
>> there would be a single SELECT_VL that decides how many scalar iterations
>> will be handled by the current vector iteration, then we would "expand"
>> the result (using MIN_EXPRs) to the multi-control cases.
For case 2 , here is the example:
+ 2. Multiple rgroup, the Gimple IR should be:
+
+ # i_23 = PHI <i_20(6), 0(11)>
+ # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
+ # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
+ # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
+ # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ _1 = i_23 * 2;
+ _2 = (long unsigned int) _1;
+ _3 = _2 * 2;
+ _4 = f_15(D) + _3;
+ _5 = _2 + 1;
+ _6 = _5 * 2;
+ _7 = f_15(D) + _6;
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ _8 = (long unsigned int) i_23;
+ _9 = _8 * 4;
+ _10 = d_18(D) + _9;
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
+ i_20 = i_23 + 1;
+ vectp_f.8_52 = vectp_f.8_56 + 16;
+ vectp_d.10_60 = vectp_d.10_63 + 16;
+ ivtmp_74 = ivtmp_73 - _75;
+ ivtmp_71 = ivtmp_70 - _72;
+ if (ivtmp_74 != 0)
+ goto <bb 6>; [83.33%]
+ else
+ goto <bb 13>; [16.67%]
+
+ Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
+ rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
+ iteration amount for mutiple rgroups.+ The analysis of the flow of multiple rgroups:
+ _72 = MIN_EXPR <ivtmp_70, 16>;
+ _75 = MIN_EXPR <ivtmp_73, 16>;
+ ...
+ .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ vectp_f.8_56 = vectp_f.8_51 + 16;
+ .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
+ ...
+ _61 = _75 / 2;
+ .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
+ vectp_d.10_63 = vectp_d.10_59 + 16;
+ _64 = _72 / 2;
+ .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);Here, If use SELECT_VL instead of MIN_EXPR. Since we define the outcome of SELECT_VL can be any number in non-final iteration.It seems not easy to adjust address pointer IV (vectp_f.8_56 = vectp_f.8_51 + 16;) and the next length (_61 = _75 / 2;).
For case 3: + 3. Multiple rgroups for non-SLP auto-vectorization.+
+ # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
+ # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
+ # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
+ _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+ _29 = (void *) ivtmp.35_10;
+ _7 = &MEM <vector([4,4]) int> [(int *)_29];
+ vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
+ _33 = _29 + POLY_INT_CST [16, 16];
+ _34 = &MEM <vector([4,4]) int> [(int *)_33];
+ vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
+ vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
+ _30 = (void *) ivtmp.36_2;
+ _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
+ .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
+ ivtmp_27 = ivtmp_26 - _28;
+ ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
+ ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
+ if (ivtmp_27 != 0)
+ goto <bb 4>; [83.33%]
+ else
+ goto <bb 5>; [16.67%]
+
+ The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
+
+ The length of first half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+
+ The length of second half vector:
+ loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
+ loop_len_16 = _28 - loop_len_15;
+
+ 1). _28 always <= POLY_INT_CST [8, 8].
+ 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
+ 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.We known in Case 3, we should deal with 2 vectors: vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;First we use _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>; to generate the number of elements to be processedfor these 2 vector.Second, we use "loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;" "loop_len_15" is the number elements to be processed for first vector.Then, "loop_len_16 = _28 - loop_len_15; "loop_len_16" is the number elements to be processed for first vector.I think "loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;" is very similiar the unpacklo in ARM SVE."loop_len_16 = _28 - loop_len_15; "loop_len_16" is very similiar the unpackhi in ARM SVE.
>> It's up to you. If you don't think select_vl is worth it then it would
>>obviously make the vectoriser changes a bit simpler.
>>But making the vectoriser simpler isn't IMO the goal here. SELECT_VL
>>seems like a perfectly reasonable construct to add to target-independent
>>code. We just need to work out some of the details.
Ok, I also prefer keeping select_vl.
>>FWIW, I share Kewen's concern about duplicating too much logic between
>>masks, current lengths, and SELECT_VL lengths. But I haven't looked at
>>the patch yet and so I don't know how easy it would be to avoid that.
I understand the concern, the current implementation are in the isolated function "vect_set_loop_controls_by_select_vl",
it's easier to review the implementation.
Maybe we can first make the whole implementation codes in "vect_set_loop_controls_by_select_vl" to be stable after review,
then we can try to incorporate these codes of "vect_set_loop_controls_by_select_vl" into "vect_set_loop_controls_directly".
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-09 20:59
To: 钟居哲
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
钟居哲 <juzhe.zhong@rivai.ai> writes:
> Hi, Richards. I would like to give more information about this patch so that it will make this patch easier for you to review.
>
> Currently, I saw we have 3 situations that we need to handle in case of loop control IV in auto-vectorization:
> 1. Single rgroup loop control (ncopies == 1 && vec_num == 1 so loop_len.length () == 1 or rgc->lengh () == 1)
> 2. Multiple rgroup for SLP.
> 3. Multiple rgroup for non-SLP which is Richard Sandiford point out previously (For example, VEC_PACK_TRUNC).
>
> To talk about this patch, let me talk about RVV LLVM implementation first which inspire me to send this patch:
> https://reviews.llvm.org/D99750
>
> According to LLVM implementation, they are adding a middle-end IR called "get_vector_length" which has totally
> same functionality as "select_vl" in this patch (I call it "while_len" previously, now I rename it as "select_vl" following Richard suggestion).
>
> The LLVM implementation is only let "get_vector_length" calculate the number of elements in single rgroup loop.
> For multi rgroup, let's take a look at it:
> https://godbolt.org/z/3GP78efTY
>
> void
> foo1 (short *__restrict f, int *__restrict d, int n)
> {
> for (int i = 0; i < n; ++i)
> {
> f[i * 2 + 0] = 1;
> f[i * 2 + 1] = 2;
> d[i] = 3;
> }
> }
>
> RISC-V Clang:
> foo1: # @foo1
> # %bb.0:
> blez a2, .LBB0_8
> # %bb.1:
> li a3, 16
> bgeu a2, a3, .LBB0_3
> # %bb.2:
> li a6, 0
> j .LBB0_6
> .LBB0_3:
> andi a6, a2, -16
> lui a3, 32
> addiw a3, a3, 1
> vsetivli zero, 8, e32, m2, ta, ma
> vmv.v.x v8, a3
> vmv.v.i v10, 3
> mv a4, a6
> mv a5, a1
> mv a3, a0
> .LBB0_4: # =>This Inner Loop Header: Depth=1
> addi a7, a5, 32
> addi t0, a3, 32
> vsetivli zero, 16, e16, m2, ta, ma
> vse16.v v8, (a3)
> vse16.v v8, (t0)
> vsetivli zero, 8, e32, m2, ta, ma
> vse32.v v10, (a5)
> vse32.v v10, (a7)
> addi a3, a3, 64
> addi a4, a4, -16
> addi a5, a5, 64
> bnez a4, .LBB0_4
> # %bb.5:
> beq a6, a2, .LBB0_8
> .LBB0_6:
> slli a3, a6, 2
> add a0, a0, a3
> addi a0, a0, 2
> add a1, a1, a3
> sub a2, a2, a6
> li a3, 1
> li a4, 2
> li a5, 3
> .LBB0_7: # =>This Inner Loop Header: Depth=1
> sh a3, -2(a0)
> sh a4, 0(a0)
> sw a5, 0(a1)
> addi a0, a0, 4
> addi a2, a2, -1
> addi a1, a1, 4
> bnez a2, .LBB0_7
> .LBB0_8:
> ret
>
> ARM GCC:
> foo1:
> cmp w2, 0
> ble .L1
> addvl x4, x0, #1
> mov x3, 0
> cntb x7
> cntb x6, all, mul #2
> sbfiz x2, x2, 1, 32
> ptrue p0.b, all
> mov x5, x2
> adrp x8, .LC0
> uqdech x5
> add x8, x8, :lo12:.LC0
> whilelo p1.h, xzr, x5
> ld1rw z1.s, p0/z, [x8]
> mov z0.s, #3
> whilelo p0.h, xzr, x2
> .L3:
> st1h z1.h, p0, [x0, x3, lsl 1]
> st1h z1.h, p1, [x4, x3, lsl 1]
> st1w z0.s, p1, [x1, #1, mul vl]
> add x3, x3, x7
> whilelo p1.h, x3, x5
> st1w z0.s, p0, [x1]
> add x1, x1, x6
> whilelo p0.h, x3, x2
> b.any .L3
> .L1:
> ret
>
> It's very obvious that ARM GCC has much better codegen since RVV LLVM just use SIMD style to handle multi-rgroup SLP auto-vectorization.
>
> Well, I am totally aggree that we should add length stuff in auto-vectorization not only for single rgroup but also multiple rgroup.
> However, when I am trying to implement multiple rgroup length for both SLP and non-SLP and testing, turns out it's hard to use select_vl
> since "select_vl" pattern allows non-VF flexible length (length <= min (remain,VF)) in any iteration, it's consuming much more operations for
> adjust loop controls IV and data reference address point IV than just using "MIN_EXPR".
>
> So for Case 2 && Case 3, I just use MIN_EXPR directly instead of SELECT_VL after my serveral internal testing.
Could you go into more details about this? I imagined that for case 3,
there would be a single SELECT_VL that decides how many scalar iterations
will be handled by the current vector iteration, then we would "expand"
the result (using MIN_EXPRs) to the multi-control cases.
In a sense that replicates what the SVE code above is doing. But for SVE,
it's possible to "optimise" the unpacking of a WHILELO result due to the
lack of implementation-defined behaviour. So conceptually we have a
single WHILELO that is unpacked one layer to give two masks. But in
reality we optimise that two WHILELOs. No such optimisation is possible
with SELECT_VL, and maybe that produces poor-quality code.
That might be what you mean (haven't had chance to look at the patch
itself yet, but hope to tomorrow).
For case 2 (max_nscalars_per_iter > 1), I think it would make conceptual
sense to pass max_nscalars_per_iter to SELECT_VL or (alternatively)
multiply the result of the SELECT_VL by max_nscalars_per_iter.
But it's only worth doing one of those two things if it produces
reasonable code for RVV.
> Now base on these situations, we only have "select_vl" for single-rgroup, but multiple-rgroup (both SLP and non-SLP), we just
> use MIN_EXPR.
>
> Is it more appropriate that we should remove "select_vl" and just use MIN_EXPR force VF elements in each non-final iteration in single rgroup?
It's up to you. If you don't think select_vl is worth it then it would
obviously make the vectoriser changes a bit simpler.
But making the vectoriser simpler isn't IMO the goal here. SELECT_VL
seems like a perfectly reasonable construct to add to target-independent
code. We just need to work out some of the details.
FWIW, I share Kewen's concern about duplicating too much logic between
masks, current lengths, and SELECT_VL lengths. But I haven't looked at
the patch yet and so I don't know how easy it would be to avoid that.
Thanks,
Richard
> Like the codegen according to RVV ISA example (show as RVV LLVM):
> https://repo.hca.bsc.es/epic/z/oynhzP
>
> ASM:
> vec_add: # @vec_add
> blez a3, .LBB0_3
> li a4, 0
> .LBB0_2: # %vector.body
> sub a5, a3, a4
> vsetvli a6, a5, e64, m1, ta, mu ==> change it into a6 = min (a5, VF) && vsetvli zero, a6, e64, m1, ta, mu
> slli a7, a4, 3
> add a5, a1, a7
> vle64.v v8, (a5)
> add a5, a2, a7
> vle64.v v9, (a5)
> vfadd.vv v8, v8, v9
> add a7, a7, a0
> add a4, a4, a6
> vse64.v v8, (a7)
> bne a4, a3, .LBB0_2
> .LBB0_3: # %for.cond.cleanup
> ret
>
> So if we remove "select_vl" just use MIN_EXPR to force VF elements in non-final iteration, we will end up with having like this:
>
> vec_add: # @vec_add
> blez a3, .LBB0_3
> csrr VF in bytes, vlenb
> .LBB0_2: # %vector.body
> sub a5, a3, VF in elements
> a6 = min (a5, VF in elements)
> vsetvli zero, a6, e64, m1, ta, mu
> add a5, a1, VF in bytes
> vle64.v v8, (a5)
> add a5, a2, VF in bytes
> vle64.v v9, (a5)
> vfadd.vv v8, v8, v9
> add a7, a7, a0
> vse64.v v8, (a7)
> bne a4, a3, .LBB0_2
> .LBB0_3: # %for.cond.cleanup
> ret
>
> Both codegen are working good for RVV.
> Just only the second one can not have the RVV special optimization (even distributed workload in last 2 iterations).
>
> Expecting any suggestions from you.
> Thanks.
>
>
> juzhe.zhong@rivai.ai
>
> From: juzhe.zhong
> Date: 2023-05-04 21:25
> To: gcc-patches
> CC: richard.sandiford; rguenther; Ju-Zhe Zhong
> Subject: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
> From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
>
> This patch is fixing V3 patch:
> https://patchwork.sourceware.org/project/gcc/patch/20230407014741.139387-1-juzhe.zhong@rivai.ai/
>
> Fix issues according to Richard Sandiford && Richard Biener.
>
> 1. Rename WHILE_LEN pattern into SELECT_VL according to Richard Sandiford.
> 2. Support multiple-rgroup for non-SLP auto-vectorization.
>
> For vec_pack_trunc pattern (multi-rgroup of non-SLP), we generate the total length:
>
> _36 = MIN_EXPR <ivtmp_34, POLY_INT_CST [8, 8]>;
>
> First length (MIN (X, VF/N)):
> loop_len_15 = MIN_EXPR <_36, POLY_INT_CST [2, 2]>;
>
> Second length (X - MIN (X, 1 * VF/N)):
> loop_len_16 = _36 - loop_len_15;
>
> Third length (X - MIN (X, 2 * VF/N)):
> _38 = MIN_EXPR <_36, POLY_INT_CST [4, 4]>;
> loop_len_17 = _36 - _38;
>
> Forth length (X - MIN (X, 3 * VF/N)):
> _39 = MIN_EXPR <_36, POLY_INT_CST [6, 6]>;
> loop_len_18 = _36 - _39;
>
> The reason that I use MIN_EXPR instead of SELECT_VL to calculate total length since using SELECT_VL
> to adapt induction IV consumes more instructions than just using MIN_EXPR. Also, during testing,
> I found it's hard to adjust length correctly according to SELECT_VL.
>
> So, this patch we only use SELECT_VL for single-rgroup with single length control.
>
> 3. Fix document of select_vl for Richard Biener (remove mode N).
> 4. Fix comments of vect_set_loop_controls_by_select_vl according to Richard Biener.
> 5. Keep loop_vinfo as first parameter for "vect_get_loop_len".
> 6. make requirement of get_while_len_data_ref_ptr outside, let it to be gated at the caller site.
>
> More comments from Richard Biener:
>>> So it's not actually saturating. The saturating operation is done by .WHILE_LEN?
> I define the outcome of SELECT_VL (n, vf) (WHILE_LEN) = IN_RANGE (0, min (n, vf)) will make
> the loop control counter never underflow zero.
>
>>> I see. I wonder if it makes sense to leave .WHILE_LEN aside for a start,
>>> the above scheme should also work for single rgroups, no?
>>> As said, it _looks_ like you can progress without .WHILE_LEN and using
>>> .WHILE_LEN is a pure optimization?
> Yes, SELECT_VL (WHILE_LEN) is pure optimization for single-rgroup and allow
> target adjust any length = INRANGE (0, min (n, vf)) each iteration.
>
> Let me known if I missed something for the V3 patch.
> Thanks.
>
> ---
> gcc/cfgloopmanip.cc | 2 +-
> gcc/doc/md.texi | 34 +++
> gcc/gimple-loop-interchange.cc | 2 +-
> gcc/internal-fn.def | 1 +
> gcc/optabs.def | 1 +
> gcc/tree-ssa-loop-ivcanon.cc | 2 +-
> gcc/tree-ssa-loop-ivopts.cc | 2 +-
> gcc/tree-ssa-loop-manip.cc | 18 +-
> gcc/tree-ssa-loop-manip.h | 4 +-
> gcc/tree-vect-data-refs.cc | 8 +-
> gcc/tree-vect-loop-manip.cc | 374 ++++++++++++++++++++++++++++++++-
> gcc/tree-vect-loop.cc | 32 ++-
> gcc/tree-vect-stmts.cc | 89 +++++++-
> gcc/tree-vectorizer.h | 4 +-
> 14 files changed, 535 insertions(+), 38 deletions(-)
>
> diff --git a/gcc/cfgloopmanip.cc b/gcc/cfgloopmanip.cc
> index 0e3ad8ed742..6e09dcbb0b1 100644
> --- a/gcc/cfgloopmanip.cc
> +++ b/gcc/cfgloopmanip.cc
> @@ -826,7 +826,7 @@ create_empty_loop_on_edge (edge entry_edge,
> }
> gsi = gsi_last_bb (loop_header);
> - create_iv (initial_value, stride, iv, loop, &gsi, false,
> + create_iv (initial_value, PLUS_EXPR, stride, iv, loop, &gsi, false,
> iv_before, iv_after);
> /* Insert loop exit condition. */
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
> index 1b77bfd46b2..e5590374e59 100644
> --- a/gcc/gimple-loop-interchange.cc
> +++ b/gcc/gimple-loop-interchange.cc
> @@ -1185,7 +1185,7 @@ tree_loop_interchange::map_inductions_to_loop (loop_cand &src, loop_cand &tgt)
> tree var_before, var_after;
> tree base = unshare_expr (iv->init_expr);
> tree step = unshare_expr (iv->step);
> - create_iv (base, step, SSA_NAME_VAR (iv->var),
> + create_iv (base, PLUS_EXPR, step, SSA_NAME_VAR (iv->var),
> tgt.m_loop, &incr_pos, false, &var_before, &var_after);
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_before));
> bitmap_set_bit (m_dce_seeds, SSA_NAME_VERSION (var_after));
> diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> index 7fe742c2ae7..6f6fa7d37f9 100644
> --- a/gcc/internal-fn.def
> +++ b/gcc/internal-fn.def
> @@ -153,6 +153,7 @@ DEF_INTERNAL_OPTAB_FN (VEC_SET, 0, vec_set, vec_set)
> DEF_INTERNAL_OPTAB_FN (LEN_STORE, 0, len_store, len_store)
> DEF_INTERNAL_OPTAB_FN (WHILE_ULT, ECF_CONST | ECF_NOTHROW, while_ult, while)
> +DEF_INTERNAL_OPTAB_FN (SELECT_VL, ECF_CONST | ECF_NOTHROW, select_vl, binary)
> DEF_INTERNAL_OPTAB_FN (CHECK_RAW_PTRS, ECF_CONST | ECF_NOTHROW,
> check_raw_ptrs, check_ptrs)
> DEF_INTERNAL_OPTAB_FN (CHECK_WAR_PTRS, ECF_CONST | ECF_NOTHROW,
> diff --git a/gcc/optabs.def b/gcc/optabs.def
> index 695f5911b30..b637471b76e 100644
> --- a/gcc/optabs.def
> +++ b/gcc/optabs.def
> @@ -476,3 +476,4 @@ OPTAB_DC (vec_series_optab, "vec_series$a", VEC_SERIES)
> OPTAB_D (vec_shl_insert_optab, "vec_shl_insert_$a")
> OPTAB_D (len_load_optab, "len_load_$a")
> OPTAB_D (len_store_optab, "len_store_$a")
> +OPTAB_D (select_vl_optab, "select_vl$a")
> diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
> index e41ec73a52a..d4e113c44df 100644
> --- a/gcc/tree-ssa-loop-ivcanon.cc
> +++ b/gcc/tree-ssa-loop-ivcanon.cc
> @@ -113,7 +113,7 @@ create_canonical_iv (class loop *loop, edge exit, tree niter,
> niter,
> build_int_cst (type, 1));
> incr_at = gsi_last_bb (in->src);
> - create_iv (niter,
> + create_iv (niter, PLUS_EXPR,
> build_int_cst (type, -1),
> NULL_TREE, loop,
> &incr_at, false, var_before, &var);
> diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
> index 78e8cbc75b5..8640fe2d487 100644
> --- a/gcc/tree-ssa-loop-ivopts.cc
> +++ b/gcc/tree-ssa-loop-ivopts.cc
> @@ -7267,7 +7267,7 @@ create_new_iv (struct ivopts_data *data, struct iv_cand *cand)
> base = unshare_expr (cand->iv->base);
> - create_iv (base, unshare_expr (cand->iv->step),
> + create_iv (base, PLUS_EXPR, unshare_expr (cand->iv->step),
> cand->var_before, data->current_loop,
> &incr_pos, after, &cand->var_before, &cand->var_after);
> }
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
> edge pe = loop_preheader_edge (loop);
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
> rewrite_all_phi_nodes_with_iv (loop, var_before);
> diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
> index d49273a3987..bda09f51d56 100644
> --- a/gcc/tree-ssa-loop-manip.h
> +++ b/gcc/tree-ssa-loop-manip.h
> @@ -22,8 +22,8 @@ along with GCC; see the file COPYING3. If not see
> typedef void (*transform_callback)(class loop *, void *);
> -extern void create_iv (tree, tree, tree, class loop *, gimple_stmt_iterator *,
> - bool, tree *, tree *);
> +extern void create_iv (tree, tree_code, tree, tree, class loop *,
> + gimple_stmt_iterator *, bool, tree *, tree *);
> extern void rewrite_into_loop_closed_ssa (bitmap, unsigned);
> extern void verify_loop_closed_ssa (bool, class loop * = NULL);
> diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
> index 6721ab6efc4..5c9103b16e5 100644
> --- a/gcc/tree-vect-data-refs.cc
> +++ b/gcc/tree-vect-data-refs.cc
> @@ -5099,7 +5099,7 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (aggr_ptr_init,
> + create_iv (aggr_ptr_init, PLUS_EXPR,
> fold_convert (aggr_ptr_type, iv_step),
> aggr_ptr, loop, &incr_gsi, insert_after,
> &indx_before_incr, &indx_after_incr);
> @@ -5129,9 +5129,9 @@ vect_create_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> {
> standard_iv_increment_position (containing_loop, &incr_gsi,
> &insert_after);
> - create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
> - containing_loop, &incr_gsi, insert_after, &indx_before_incr,
> - &indx_after_incr);
> + create_iv (aptr, PLUS_EXPR, fold_convert (aggr_ptr_type, DR_STEP (dr)),
> + aggr_ptr, containing_loop, &incr_gsi, insert_after,
> + &indx_before_incr, &indx_after_incr);
> incr = gsi_stmt (incr_gsi);
> /* Copy the points-to information if it exists. */
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> @@ -468,8 +510,9 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (build_int_cst (iv_type, 0), nitems_step, NULL_TREE, loop,
> - &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
> + create_iv (build_int_cst (iv_type, 0), PLUS_EXPR, nitems_step, NULL_TREE,
> + loop, &incr_gsi, insert_after, &index_before_incr,
> + &index_after_incr);
> tree zero_index = build_int_cst (compare_type, 0);
> tree test_index, test_limit, first_limit;
> @@ -682,6 +725,300 @@ vect_set_loop_controls_directly (class loop *loop, loop_vec_info loop_vinfo,
> return next_ctrl;
> }
> +/* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> + for all the rgroup controls in RGC and return a control that is nonzero
> + when the loop needs to iterate. Add any new preheader statements to
> + PREHEADER_SEQ. Use LOOP_COND_GSI to insert code before the exit gcond.
> +
> + RGC belongs to loop LOOP. The loop originally iterated NITERS
> + times and has been vectorized according to LOOP_VINFO.
> +
> + Unlike vect_set_loop_controls_directly which is iterating from 0-based IV
> + to TEST_LIMIT - bias.
> +
> + In vect_set_loop_controls_by_select_vl, we are iterating from start at
> + IV = TEST_LIMIT - bias and keep subtract IV by the length calculated by
> + IFN_SELECT_VL pattern.
> +
> + 1. Single rgroup, the Gimple IR should be:
> +
> + # vectp_B.6_8 = PHI <vectp_B.6_13(6), &B(5)>
> + # vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
> + # vectp_A.11_19 = PHI <vectp_A.11_20(6), &A(5)>
> + # vectp_A.13_22 = PHI <vectp_A.13_23(6), &A(5)>
> + # ivtmp_26 = PHI <ivtmp_27(6), _25(5)>
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + ivtmp_15 = _28 * 4;
> + vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);
> + _1 = B[i_10];
> + .LEN_STORE (vectp_A.13_22, 128B, _28, vect__1.10_18, 0);
> + i_7 = i_10 + 1;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + Note: We use the outcome of .SELECT_VL to adjust both loop control IV and
> + data reference pointer IV.
> +
> + 1). The result of .SELECT_VL:
> + _28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);
> + The _28 is not necessary to be VF in any iteration, instead, we allow
> + _28 to be any value as long as _28 <= VF. Such flexible SELECT_VL
> + pattern allows target have various flexible optimizations in vector
> + loop iterations. Target like RISC-V has special application vector
> + length calculation instruction which will distribute even workload
> + in the last 2 iterations.
> +
> + Other example is that we can allow even generate _28 <= VF / 2 so
> + that some machine can run vector codes in low power mode.
> +
> + 2). Loop control IV:
> + ivtmp_27 = ivtmp_26 - _28;
> + if (ivtmp_27 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 7>; [16.67%]
> +
> + This is the saturating-subtraction towards zero, the outcome of
> + .SELECT_VL wil make ivtmp_27 never underflow zero.
> +
> + 3). Data reference pointer IV:
> + ivtmp_15 = _28 * 4;
> + vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
> + vectp_A.13_23 = vectp_A.13_22 + ivtmp_15;
> +
> + The pointer IV is adjusted accurately according to the .SELECT_VL.
> +
> + 2. Multiple rgroup, the Gimple IR should be:
> +
> + # i_23 = PHI <i_20(6), 0(11)>
> + # vectp_f.8_51 = PHI <vectp_f.8_52(6), f_15(D)(11)>
> + # vectp_d.10_59 = PHI <vectp_d.10_60(6), d_18(D)(11)>
> + # ivtmp_70 = PHI <ivtmp_71(6), _69(11)>
> + # ivtmp_73 = PHI <ivtmp_74(6), _67(11)>
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + _1 = i_23 * 2;
> + _2 = (long unsigned int) _1;
> + _3 = _2 * 2;
> + _4 = f_15(D) + _3;
> + _5 = _2 + 1;
> + _6 = _5 * 2;
> + _7 = f_15(D) + _6;
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _8 = (long unsigned int) i_23;
> + _9 = _8 * 4;
> + _10 = d_18(D) + _9;
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + i_20 = i_23 + 1;
> + vectp_f.8_52 = vectp_f.8_56 + 16;
> + vectp_d.10_60 = vectp_d.10_63 + 16;
> + ivtmp_74 = ivtmp_73 - _75;
> + ivtmp_71 = ivtmp_70 - _72;
> + if (ivtmp_74 != 0)
> + goto <bb 6>; [83.33%]
> + else
> + goto <bb 13>; [16.67%]
> +
> + Note: We DO NOT use .SELECT_VL in SLP auto-vectorization for multiple
> + rgroups. Instead, we use MIN_EXPR to guarantee we always use VF as the
> + iteration amount for mutiple rgroups.
> +
> + The analysis of the flow of multiple rgroups:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + _75 = MIN_EXPR <ivtmp_73, 16>;
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _75, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + vectp_f.8_56 = vectp_f.8_51 + 16;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + ...
> + _61 = _75 / 2;
> + .LEN_STORE (vectp_d.10_59, 128B, _61, { 3, 3, 3, 3 }, 0);
> + vectp_d.10_63 = vectp_d.10_59 + 16;
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> +
> + We use _72 = MIN_EXPR <ivtmp_70, 16>; to generate the number of the elements
> + to be processed in each iteration.
> +
> + The related STOREs:
> + _72 = MIN_EXPR <ivtmp_70, 16>;
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2, 1, 2, 1, 2 }, 0);
> + _64 = _72 / 2;
> + .LEN_STORE (vectp_d.10_63, 128B, _64, { 3, 3, 3, 3 }, 0);
> + Since these 2 STOREs store 2 vectors that the second vector is half elements
> + of the first vector. So the length of second STORE will be _64 = _72 / 2;
> + It's similar to the VIEW_CONVERT of handling masks in SLP.
> +
> + 3. Multiple rgroups for non-SLP auto-vectorization.
> +
> + # ivtmp_26 = PHI <ivtmp_27(4), _25(3)>
> + # ivtmp.35_10 = PHI <ivtmp.35_11(4), ivtmp.35_1(3)>
> + # ivtmp.36_2 = PHI <ivtmp.36_8(4), ivtmp.36_23(3)>
> + _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> + _29 = (void *) ivtmp.35_10;
> + _7 = &MEM <vector([4,4]) int> [(int *)_29];
> + vect__1.25_17 = .LEN_LOAD (_7, 128B, loop_len_15, 0);
> + _33 = _29 + POLY_INT_CST [16, 16];
> + _34 = &MEM <vector([4,4]) int> [(int *)_33];
> + vect__1.26_19 = .LEN_LOAD (_34, 128B, loop_len_16, 0);
> + vect__2.27_20 = VEC_PACK_TRUNC_EXPR <vect__1.25_17, vect__1.26_19>;
> + _30 = (void *) ivtmp.36_2;
> + _31 = &MEM <vector([8,8]) short int> [(short int *)_30];
> + .LEN_STORE (_31, 128B, _28, vect__2.27_20, 0);
> + ivtmp_27 = ivtmp_26 - _28;
> + ivtmp.35_11 = ivtmp.35_10 + POLY_INT_CST [32, 32];
> + ivtmp.36_8 = ivtmp.36_2 + POLY_INT_CST [16, 16];
> + if (ivtmp_27 != 0)
> + goto <bb 4>; [83.33%]
> + else
> + goto <bb 5>; [16.67%]
> +
> + The total length: _28 = MIN_EXPR <ivtmp_26, POLY_INT_CST [8, 8]>;
> +
> + The length of first half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> +
> + The length of second half vector:
> + loop_len_15 = MIN_EXPR <_28, POLY_INT_CST [4, 4]>;
> + loop_len_16 = _28 - loop_len_15;
> +
> + 1). _28 always <= POLY_INT_CST [8, 8].
> + 2). When _28 <= POLY_INT_CST [4, 4], second half vector is not processed.
> + 3). When _28 > POLY_INT_CST [4, 4], second half vector is processed.
> +*/
> +
> +static tree
> +vect_set_loop_controls_by_select_vl (class loop *loop, loop_vec_info loop_vinfo,
> + gimple_seq *preheader_seq,
> + gimple_seq *header_seq,
> + rgroup_controls *rgc, tree niters,
> + unsigned int controls_length)
> +{
> + tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + /* We are not allowing masked approach in SELECT_VL. */
> + gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo));
> +
> + tree ctrl_type = rgc->type;
> + unsigned int nitems_per_iter = rgc->max_nscalars_per_iter * rgc->factor;
> + poly_uint64 nitems_per_ctrl = TYPE_VECTOR_SUBPARTS (ctrl_type) * rgc->factor;
> + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
> +
> + /* Calculate the maximum number of item values that the rgroup
> + handles in total, the number that it handles for each iteration
> + of the vector loop. */
> + tree nitems_total = niters;
> + if (nitems_per_iter != 1)
> + {
> + /* We checked before setting LOOP_VINFO_USING_PARTIAL_VECTORS_P that
> + these multiplications don't overflow. */
> + tree compare_factor = build_int_cst (compare_type, nitems_per_iter);
> + nitems_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
> + nitems_total, compare_factor);
> + }
> +
> + /* Convert the comparison value to the IV type (either a no-op or
> + a promotion). */
> + nitems_total = gimple_convert (preheader_seq, iv_type, nitems_total);
> +
> + /* Create an induction variable that counts the number of items
> + processed. */
> + tree index_before_incr, index_after_incr;
> + gimple_stmt_iterator incr_gsi;
> + bool insert_after;
> + standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> +
> + /* Test the decremented IV, which will never underflow 0 since we have
> + IFN_SELECT_VL to gurantee that. */
> + tree test_limit = nitems_total;
> +
> + /* Provide a definition of each control in the group. */
> + tree ctrl;
> + unsigned int i;
> + FOR_EACH_VEC_ELT_REVERSE (rgc->controls, i, ctrl)
> + {
> + /* Previous controls will cover BIAS items. This control covers the
> + next batch. */
> + poly_uint64 bias = nitems_per_ctrl * i;
> + tree bias_tree = build_int_cst (iv_type, bias);
> +
> + /* Rather than have a new IV that starts at TEST_LIMIT and goes down to
> + BIAS, prefer to use the same TEST_LIMIT - BIAS based IV for each
> + control and adjust the bound down by BIAS. */
> + tree this_test_limit = test_limit;
> + if (i != 0)
> + {
> + this_test_limit = gimple_build (preheader_seq, MAX_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, iv_type,
> + this_test_limit, bias_tree);
> + }
> +
> + /* Create decrement IV. */
> + create_iv (this_test_limit, MINUS_EXPR, ctrl, NULL_TREE, loop, &incr_gsi,
> + insert_after, &index_before_incr, &index_after_incr);
> +
> + poly_uint64 final_vf = vf * nitems_per_iter;
> + tree vf_step = build_int_cst (iv_type, final_vf);
> + tree res_len;
> + if (nitems_per_iter != 1 || controls_length != 1)
> + {
> + /* For SLP, we can't allow non-VF number of elements to be processed
> + in non-final iteration. We force the number of elements to be
> + processed in each non-final iteration is VF elements. If we allow
> + non-VF elements processing in non-final iteration will make SLP too
> + complicated and produce inferior codegen.
> +
> + For example:
> +
> + If non-final iteration process VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + ...
> +
> + If non-final iteration process non-VF elements.
> +
> + ...
> + .LEN_STORE (vectp_f.8_51, 128B, _71, { 1, 2, 1, 2 }, 0);
> + if (_71 % 2 == 0)
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 1, 2, 1, 2 }, 0);
> + else
> + .LEN_STORE (vectp_f.8_56, 128B, _72, { 2, 1, 2, 1 }, 0);
> + ...
> +
> + This is the simple case of 2-elements interleaved vector SLP. We
> + consider other interleave vector, the situation will become more
> + complicated. */
> + res_len = gimple_build (header_seq, MIN_EXPR, iv_type,
> + index_before_incr, vf_step);
> + }
> + else
> + {
> + res_len = gimple_build (header_seq, IFN_SELECT_VL, iv_type,
> + index_before_incr, vf_step);
> + }
> + gassign *assign = gimple_build_assign (ctrl, res_len);
> + gimple_seq_add_stmt (header_seq, assign);
> + }
> +
> + return index_after_incr;
> +}
> +
> /* Set up the iteration condition and rgroup controls for LOOP, given
> that LOOP_VINFO_USING_PARTIAL_VECTORS_P is true for the vectorized
> loop. LOOP_VINFO describes the vectorization of LOOP. NITERS is
> @@ -703,6 +1040,10 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> bool use_masks_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> tree compare_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo);
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> + bool use_vl_p = !use_masks_p
> + && direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED);
> unsigned int compare_precision = TYPE_PRECISION (compare_type);
> tree orig_niters = niters;
> @@ -752,17 +1093,32 @@ vect_set_loop_condition_partial_vectors (class loop *loop,
> continue;
> }
> + if (use_vl_p && rgc->controls.length () != 1)
> + {
> + rgroup_controls *sub_rgc
> + = &(*controls)[nmasks / rgc->controls.length () - 1];
> + if (!sub_rgc->controls.is_empty ()
> + && vect_maybe_permute_loop_lens (iv_type, &header_seq, rgc,
> + sub_rgc))
> + continue;
> + }
> +
> /* See whether zero-based IV would ever generate all-false masks
> or zero length before wrapping around. */
> bool might_wrap_p = vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc);
> /* Set up all controls for this group. */
> - test_ctrl = vect_set_loop_controls_directly (loop, loop_vinfo,
> - &preheader_seq,
> - &header_seq,
> - loop_cond_gsi, rgc,
> - niters, niters_skip,
> - might_wrap_p);
> + if (use_vl_p)
> + test_ctrl
> + = vect_set_loop_controls_by_select_vl (loop, loop_vinfo,
> + &preheader_seq, &header_seq,
> + rgc, niters, controls->length ());
> + else
> + test_ctrl
> + = vect_set_loop_controls_directly (loop, loop_vinfo, &preheader_seq,
> + &header_seq, loop_cond_gsi, rgc,
> + niters, niters_skip,
> + might_wrap_p);
> }
> /* Emit all accumulated statements. */
> @@ -893,7 +1249,7 @@ vect_set_loop_condition_normal (class loop *loop, tree niters, tree step,
> }
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (init, step, NULL_TREE, loop,
> + create_iv (init, PLUS_EXPR, step, NULL_TREE, loop,
> &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
> indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
> true, NULL_TREE, true,
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 6ea0f21fd13..d7cf92576d1 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -5567,7 +5567,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
> gimple_stmt_iterator incr_gsi;
> bool insert_after;
> standard_iv_increment_position (loop, &incr_gsi, &insert_after);
> - create_iv (series_vect, vec_step, NULL_TREE, loop, &incr_gsi,
> + create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi,
> insert_after, &indx_before_incr, &indx_after_incr);
> /* Next create a new phi node vector (NEW_PHI_TREE) which starts
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-04 13:25 [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support juzhe.zhong
` (2 preceding siblings ...)
2023-05-08 5:35 ` Kewen.Lin
@ 2023-05-10 16:45 ` Richard Sandiford
2023-05-10 21:00 ` 钟居哲
2023-05-11 10:11 ` juzhe.zhong
3 siblings, 2 replies; 25+ messages in thread
From: Richard Sandiford @ 2023-05-10 16:45 UTC (permalink / raw)
To: juzhe.zhong; +Cc: gcc-patches, rguenther
In addition to Jeff's comments:
juzhe.zhong@rivai.ai writes:
> [...]
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
GCC's IN_RANGE is a predicate, so it would be best to avoid that here.
Why isn't it simply ceil (operand1 / 2), which must be <= operand2?
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> [...]
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
>
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
As Richard said, we should be able to get rid of incr_op, probably
by calling the parameter incr_op.
I think the later:
incr_op = MINUS_EXPR;
stmts need to be updated to something that flips between PLUS_EXPR
and MINUS_EXPR (with updates to the comments).
It would probably make sense to split the create_iv part out as a
separate prepatch.
> edge pe = loop_preheader_edge (loop);
>
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
>
> rewrite_all_phi_nodes_with_iv (loop, var_before);
>
> [...]
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
Nit: think it's better to talk about "lengths" rather than "lens"
(which ordinarily means something different).
Like Jeff said, this operation is no longer a permute when applied
to lengths (as opposed to masks). And because of that, I don't think
there's anything special about DEST_RGN having twice as many controls
as SRC_RGN. (And the implementation does seem to be more general
than just a factor of two.)
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
This return false thing looks odd. I don't think we can decide on
a group-by-group basis whether to use this approach or not. Instead
it has to be a per-loop decision. Either (1) we use SELECT_VL for the
single-control group, and use this function to derive all other groups
from that, or (2) we just MIN for all groups. No mixing and matching
is allowed.
In other words, I think we need to decide at the top level whether
to use SELECT_VL or not. As things stand, this decision should be
"no" if nitems_per_iter != 1 for *any* group. If the decision is
to use SELECT_VL, we should use it for the single-control group
only and then unconditionally use the function above for the other
groups, with no failure path.
The decision about whether to use SELECT_VL or not should be cached
if necessary, so that it's available...
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> [...]
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
>
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
>
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
>
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
>
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
...here. That is, the key isn't whether SELECT_VL is available,
but instead whether we've decided to use it for this loop (unless
I'm missing something).
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
>
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
It seems a little wasteful to recompute the temporary results (into gsi)
for every pointer, but I guess it's not worth complicating the code to
do on-the-fly DCE. Later passes can clean it up for us. (Unless Richi
feels differently. Can't remember who's said what now.)
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
>
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
Here too I think the check should be based on a cached record
of whether we're using SELECT_VL for this particular loop,
What happens for loop_lens->length () != 1?
Thanks,
Richard
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
>
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-10 16:45 ` Richard Sandiford
@ 2023-05-10 21:00 ` 钟居哲
2023-05-10 21:28 ` Richard Sandiford
2023-05-11 10:11 ` juzhe.zhong
1 sibling, 1 reply; 25+ messages in thread
From: 钟居哲 @ 2023-05-10 21:00 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 20665 bytes --]
Thanks Richard.
I am planning to seperate a patch with only creat_iv stuff only.
Are you suggesting that I remove "tree_code incr_op = code;"
Use the argument directly ?
I saw the codes here:
/* For easier readability of the created code, produce MINUS_EXPRs
when suitable. */
if (TREE_CODE (step) == INTEGER_CST)
{
if (TYPE_UNSIGNED (TREE_TYPE (step)))
{
step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
if (tree_int_cst_lt (step1, step))
{
incr_op = MINUS_EXPR;
step = step1;
}
}
else
{
bool ovf;
if (!tree_expr_nonnegative_warnv_p (step, &ovf)
&& may_negate_without_overflow_p (step))
{
incr_op = MINUS_EXPR;
step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
}
}
}
if (POINTER_TYPE_P (TREE_TYPE (base)))
{
if (TREE_CODE (base) == ADDR_EXPR)
mark_addressable (TREE_OPERAND (base, 0));
step = convert_to_ptrofftype (step);
if (incr_op == MINUS_EXPR)
step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
incr_op = POINTER_PLUS_EXPR;
}
/* Gimplify the step if necessary. We put the computations in front of the
loop (i.e. the step should be loop invariant). */
step = force_gimple_operand (step, &stmts, true, NULL_TREE);
if (stmts)
gsi_insert_seq_on_edge_immediate (pe, stmts);
stmt = gimple_build_assign (va, incr_op, vb, step);
...
It seems that it has complicated conditions here to change value of variable "incr_op".
That's why I define a temporary variable "tree_code incr_op = code;" here and
let the following codes change the value of "incr_op".
Could you give me some hints of dealing with this piece of code to get rid of "tree_code incr_op = code;" ?
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 00:45
To: juzhe.zhong
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
In addition to Jeff's comments:
juzhe.zhong@rivai.ai writes:
> [...]
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
GCC's IN_RANGE is a predicate, so it would be best to avoid that here.
Why isn't it simply ceil (operand1 / 2), which must be <= operand2?
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> [...]
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
>
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
As Richard said, we should be able to get rid of incr_op, probably
by calling the parameter incr_op.
I think the later:
incr_op = MINUS_EXPR;
stmts need to be updated to something that flips between PLUS_EXPR
and MINUS_EXPR (with updates to the comments).
It would probably make sense to split the create_iv part out as a
separate prepatch.
> edge pe = loop_preheader_edge (loop);
>
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
>
> rewrite_all_phi_nodes_with_iv (loop, var_before);
>
> [...]
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
Nit: think it's better to talk about "lengths" rather than "lens"
(which ordinarily means something different).
Like Jeff said, this operation is no longer a permute when applied
to lengths (as opposed to masks). And because of that, I don't think
there's anything special about DEST_RGN having twice as many controls
as SRC_RGN. (And the implementation does seem to be more general
than just a factor of two.)
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
This return false thing looks odd. I don't think we can decide on
a group-by-group basis whether to use this approach or not. Instead
it has to be a per-loop decision. Either (1) we use SELECT_VL for the
single-control group, and use this function to derive all other groups
from that, or (2) we just MIN for all groups. No mixing and matching
is allowed.
In other words, I think we need to decide at the top level whether
to use SELECT_VL or not. As things stand, this decision should be
"no" if nitems_per_iter != 1 for *any* group. If the decision is
to use SELECT_VL, we should use it for the single-control group
only and then unconditionally use the function above for the other
groups, with no failure path.
The decision about whether to use SELECT_VL or not should be cached
if necessary, so that it's available...
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> [...]
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
>
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
>
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
>
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
>
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
...here. That is, the key isn't whether SELECT_VL is available,
but instead whether we've decided to use it for this loop (unless
I'm missing something).
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
>
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
It seems a little wasteful to recompute the temporary results (into gsi)
for every pointer, but I guess it's not worth complicating the code to
do on-the-fly DCE. Later passes can clean it up for us. (Unless Richi
feels differently. Can't remember who's said what now.)
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
>
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
Here too I think the check should be based on a cached record
of whether we're using SELECT_VL for this particular loop,
What happens for loop_lens->length () != 1?
Thanks,
Richard
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
>
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-10 21:00 ` 钟居哲
@ 2023-05-10 21:28 ` Richard Sandiford
2023-05-10 22:51 ` 钟居哲
0 siblings, 1 reply; 25+ messages in thread
From: Richard Sandiford @ 2023-05-10 21:28 UTC (permalink / raw)
To: 钟居哲; +Cc: gcc-patches, rguenther
钟居哲 <juzhe.zhong@rivai.ai> writes:
> Thanks Richard.
> I am planning to seperate a patch with only creat_iv stuff only.
>
> Are you suggesting that I remove "tree_code incr_op = code;"
> Use the argument directly ?
>
> I saw the codes here:
>
> /* For easier readability of the created code, produce MINUS_EXPRs
> when suitable. */
> if (TREE_CODE (step) == INTEGER_CST)
> {
> if (TYPE_UNSIGNED (TREE_TYPE (step)))
> {
> step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> if (tree_int_cst_lt (step1, step))
> {
> incr_op = MINUS_EXPR;
> step = step1;
> }
> }
> else
> {
> bool ovf;
>
> if (!tree_expr_nonnegative_warnv_p (step, &ovf)
> && may_negate_without_overflow_p (step))
> {
> incr_op = MINUS_EXPR;
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> }
> }
> }
> if (POINTER_TYPE_P (TREE_TYPE (base)))
> {
> if (TREE_CODE (base) == ADDR_EXPR)
> mark_addressable (TREE_OPERAND (base, 0));
> step = convert_to_ptrofftype (step);
> if (incr_op == MINUS_EXPR)
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> incr_op = POINTER_PLUS_EXPR;
> }
> /* Gimplify the step if necessary. We put the computations in front of the
> loop (i.e. the step should be loop invariant). */
> step = force_gimple_operand (step, &stmts, true, NULL_TREE);
> if (stmts)
> gsi_insert_seq_on_edge_immediate (pe, stmts);
>
> stmt = gimple_build_assign (va, incr_op, vb, step);
> ...
>
> It seems that it has complicated conditions here to change value of variable "incr_op".
> That's why I define a temporary variable "tree_code incr_op = code;" here and
> let the following codes change the value of "incr_op".
>
> Could you give me some hints of dealing with this piece of code to get rid of "tree_code incr_op = code;" ?
Yeah, but like I said in the review, those later:
incr_op = MINUS_EXPR;
stmts need to be updated to something that flips between PLUS_EXPR
and MINUS_EXPR (with updates to the comments). Just leaving them
as-is is incorrect (in cases where the caller passed MINUS_EXPR
rather than PLUS_EXPR).
The POINTER_PLUS_EXPR handling is fine due to the conditional
negate beforehand.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-10 21:28 ` Richard Sandiford
@ 2023-05-10 22:51 ` 钟居哲
2023-05-11 4:50 ` Richard Sandiford
0 siblings, 1 reply; 25+ messages in thread
From: 钟居哲 @ 2023-05-10 22:51 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 4560 bytes --]
I am sorry that I am still confused about that.
Is this what you want ?
bool use_minus_p = TREE_CODE (step) == INTEGER_CST && ((TYPE_UNSIGNED (TREE_TYPE (step)) && tree_int_cst_lt (step1, step))
|| (!TYPE_UNSIGNED (TREE_TYPE (step)) && !tree_expr_nonnegative_warnv_p (step, &ovf) && may_negate_without_overflow_p (step)));
/* For easier readability of the created code, produce MINUS_EXPRs
when suitable. */
if (TREE_CODE (step) == INTEGER_CST)
{
if (TYPE_UNSIGNED (TREE_TYPE (step)))
{
step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
if (tree_int_cst_lt (step1, step))
{
incr_op = MINUS_EXPR; /* Remove it. */
step = step1;
}
}
else
{
bool ovf;
if (!tree_expr_nonnegative_warnv_p (step, &ovf)
&& may_negate_without_overflow_p (step))
{
incr_op = MINUS_EXPR; /* Remove it. */
step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
}
}
}
if (POINTER_TYPE_P (TREE_TYPE (base)))
{
if (TREE_CODE (base) == ADDR_EXPR)
mark_addressable (TREE_OPERAND (base, 0));
step = convert_to_ptrofftype (step);
if (incr_op == MINUS_EXPR) /* Change it into if (use_minus_p) */
step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
incr_op = POINTER_PLUS_EXPR; /* Remove it. */
}
/* Gimplify the step if necessary. We put the computations in front of the
loop (i.e. the step should be loop invariant). */
step = force_gimple_operand (step, &stmts, true, NULL_TREE);
if (stmts)
gsi_insert_seq_on_edge_immediate (pe, stmts);
if (POINTER_TYPE_P (TREE_TYPE (base)))
stmt = gimple_build_assign (va, POINTER_PLUS_EXPR, vb, step);
else if (use_minus_p)
stmt = gimple_build_assign (va, MINUS_EXPR, vb, step);
else
stmt = gimple_build_assign (va, incr_op, vb, step);
...
Since I have no idea to make stmts flips between PLUS_EXPR and MINUS_EXPR.
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 05:28
To: 钟居哲
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
钟居哲 <juzhe.zhong@rivai.ai> writes:
> Thanks Richard.
> I am planning to seperate a patch with only creat_iv stuff only.
>
> Are you suggesting that I remove "tree_code incr_op = code;"
> Use the argument directly ?
>
> I saw the codes here:
>
> /* For easier readability of the created code, produce MINUS_EXPRs
> when suitable. */
> if (TREE_CODE (step) == INTEGER_CST)
> {
> if (TYPE_UNSIGNED (TREE_TYPE (step)))
> {
> step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> if (tree_int_cst_lt (step1, step))
> {
> incr_op = MINUS_EXPR;
> step = step1;
> }
> }
> else
> {
> bool ovf;
>
> if (!tree_expr_nonnegative_warnv_p (step, &ovf)
> && may_negate_without_overflow_p (step))
> {
> incr_op = MINUS_EXPR;
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> }
> }
> }
> if (POINTER_TYPE_P (TREE_TYPE (base)))
> {
> if (TREE_CODE (base) == ADDR_EXPR)
> mark_addressable (TREE_OPERAND (base, 0));
> step = convert_to_ptrofftype (step);
> if (incr_op == MINUS_EXPR)
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> incr_op = POINTER_PLUS_EXPR;
> }
> /* Gimplify the step if necessary. We put the computations in front of the
> loop (i.e. the step should be loop invariant). */
> step = force_gimple_operand (step, &stmts, true, NULL_TREE);
> if (stmts)
> gsi_insert_seq_on_edge_immediate (pe, stmts);
>
> stmt = gimple_build_assign (va, incr_op, vb, step);
> ...
>
> It seems that it has complicated conditions here to change value of variable "incr_op".
> That's why I define a temporary variable "tree_code incr_op = code;" here and
> let the following codes change the value of "incr_op".
>
> Could you give me some hints of dealing with this piece of code to get rid of "tree_code incr_op = code;" ?
Yeah, but like I said in the review, those later:
incr_op = MINUS_EXPR;
stmts need to be updated to something that flips between PLUS_EXPR
and MINUS_EXPR (with updates to the comments). Just leaving them
as-is is incorrect (in cases where the caller passed MINUS_EXPR
rather than PLUS_EXPR).
The POINTER_PLUS_EXPR handling is fine due to the conditional
negate beforehand.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-10 22:51 ` 钟居哲
@ 2023-05-11 4:50 ` Richard Sandiford
2023-05-11 5:14 ` juzhe.zhong
0 siblings, 1 reply; 25+ messages in thread
From: Richard Sandiford @ 2023-05-11 4:50 UTC (permalink / raw)
To: 钟居哲; +Cc: gcc-patches, rguenther
钟居哲 <juzhe.zhong@rivai.ai> writes:
> I am sorry that I am still confused about that.
>
> Is this what you want ?
>
> bool use_minus_p = TREE_CODE (step) == INTEGER_CST && ((TYPE_UNSIGNED (TREE_TYPE (step)) && tree_int_cst_lt (step1, step))
> || (!TYPE_UNSIGNED (TREE_TYPE (step)) && !tree_expr_nonnegative_warnv_p (step, &ovf) && may_negate_without_overflow_p (step)));
>
> /* For easier readability of the created code, produce MINUS_EXPRs
> when suitable. */
> if (TREE_CODE (step) == INTEGER_CST)
> {
> if (TYPE_UNSIGNED (TREE_TYPE (step)))
> {
> step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> if (tree_int_cst_lt (step1, step))
> {
> incr_op = MINUS_EXPR; /* Remove it. */
> step = step1;
> }
> }
> else
> {
> bool ovf;
>
> if (!tree_expr_nonnegative_warnv_p (step, &ovf)
> && may_negate_without_overflow_p (step))
> {
> incr_op = MINUS_EXPR; /* Remove it. */
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> }
> }
> }
> if (POINTER_TYPE_P (TREE_TYPE (base)))
> {
> if (TREE_CODE (base) == ADDR_EXPR)
> mark_addressable (TREE_OPERAND (base, 0));
> step = convert_to_ptrofftype (step);
> if (incr_op == MINUS_EXPR) /* Change it into if (use_minus_p) */
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> incr_op = POINTER_PLUS_EXPR; /* Remove it. */
> }
> /* Gimplify the step if necessary. We put the computations in front of the
> loop (i.e. the step should be loop invariant). */
> step = force_gimple_operand (step, &stmts, true, NULL_TREE);
> if (stmts)
> gsi_insert_seq_on_edge_immediate (pe, stmts);
>
> if (POINTER_TYPE_P (TREE_TYPE (base)))
> stmt = gimple_build_assign (va, POINTER_PLUS_EXPR, vb, step);
> else if (use_minus_p)
> stmt = gimple_build_assign (va, MINUS_EXPR, vb, step);
> else
> stmt = gimple_build_assign (va, incr_op, vb, step);
> ...
>
> Since I have no idea to make stmts flips between PLUS_EXPR and MINUS_EXPR.
No, I meant:
- Rename the "code" argument to "incr_op".
- Remove "tree_code incr_op = code;".
- Replace both instances of:
incr_op = MINUS_EXPR;
with:
incr_op = (incr_op == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR);
The point is that the current code (rightly) assumes that incr_op
always starts out as PLUS_EXPR, i.e. that STEP starts out applying
positively. Making STEP apply in the opposite direction is then as
simple as changing incr_op to MINUS_EXPR. But the new interface
allows STEP to start out applying positively or negatively, and so
this code needs to cope with both cases.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 4:50 ` Richard Sandiford
@ 2023-05-11 5:14 ` juzhe.zhong
0 siblings, 0 replies; 25+ messages in thread
From: juzhe.zhong @ 2023-05-11 5:14 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 3136 bytes --]
Thank you so much.
Can you take a look at this patch:
https://gcc.gnu.org/pipermail/gcc-patches/2023-May/618110.html
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 12:50
To: 钟居哲
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
钟居哲 <juzhe.zhong@rivai.ai> writes:
> I am sorry that I am still confused about that.
>
> Is this what you want ?
>
> bool use_minus_p = TREE_CODE (step) == INTEGER_CST && ((TYPE_UNSIGNED (TREE_TYPE (step)) && tree_int_cst_lt (step1, step))
> || (!TYPE_UNSIGNED (TREE_TYPE (step)) && !tree_expr_nonnegative_warnv_p (step, &ovf) && may_negate_without_overflow_p (step)));
>
> /* For easier readability of the created code, produce MINUS_EXPRs
> when suitable. */
> if (TREE_CODE (step) == INTEGER_CST)
> {
> if (TYPE_UNSIGNED (TREE_TYPE (step)))
> {
> step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> if (tree_int_cst_lt (step1, step))
> {
> incr_op = MINUS_EXPR; /* Remove it. */
> step = step1;
> }
> }
> else
> {
> bool ovf;
>
> if (!tree_expr_nonnegative_warnv_p (step, &ovf)
> && may_negate_without_overflow_p (step))
> {
> incr_op = MINUS_EXPR; /* Remove it. */
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> }
> }
> }
> if (POINTER_TYPE_P (TREE_TYPE (base)))
> {
> if (TREE_CODE (base) == ADDR_EXPR)
> mark_addressable (TREE_OPERAND (base, 0));
> step = convert_to_ptrofftype (step);
> if (incr_op == MINUS_EXPR) /* Change it into if (use_minus_p) */
> step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
> incr_op = POINTER_PLUS_EXPR; /* Remove it. */
> }
> /* Gimplify the step if necessary. We put the computations in front of the
> loop (i.e. the step should be loop invariant). */
> step = force_gimple_operand (step, &stmts, true, NULL_TREE);
> if (stmts)
> gsi_insert_seq_on_edge_immediate (pe, stmts);
>
> if (POINTER_TYPE_P (TREE_TYPE (base)))
> stmt = gimple_build_assign (va, POINTER_PLUS_EXPR, vb, step);
> else if (use_minus_p)
> stmt = gimple_build_assign (va, MINUS_EXPR, vb, step);
> else
> stmt = gimple_build_assign (va, incr_op, vb, step);
> ...
>
> Since I have no idea to make stmts flips between PLUS_EXPR and MINUS_EXPR.
No, I meant:
- Rename the "code" argument to "incr_op".
- Remove "tree_code incr_op = code;".
- Replace both instances of:
incr_op = MINUS_EXPR;
with:
incr_op = (incr_op == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR);
The point is that the current code (rightly) assumes that incr_op
always starts out as PLUS_EXPR, i.e. that STEP starts out applying
positively. Making STEP apply in the opposite direction is then as
simple as changing incr_op to MINUS_EXPR. But the new interface
allows STEP to start out applying positively or negatively, and so
this code needs to cope with both cases.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-10 16:45 ` Richard Sandiford
2023-05-10 21:00 ` 钟居哲
@ 2023-05-11 10:11 ` juzhe.zhong
2023-05-11 11:04 ` Richard Sandiford
1 sibling, 1 reply; 25+ messages in thread
From: juzhe.zhong @ 2023-05-11 10:11 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 22405 bytes --]
Hi， Richard. Since create_iv has been approved and soon will be commited after
we bootstrap && regression.
Now, I plan to send patch for "decrement IV".
After reading your comments, I have several questions:
1.
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
> ...here. That is, the key isn't whether SELECT_VL is available,
> but instead whether we've decided to use it for this loop (unless
> I'm missing something).
Let's me clarify it again:
I do this here is for Case 2 SLP:
Generate for len : _61 = _75 / 2;
I think it is similar with ARM SVE using VIEW_CONVER_EXPR to view_convert the mask.
You said we should not let SELECT_VL is available or not to decide it here.
Could you teach me how to handle this code here? Should I add a target hook like:
TARGET_SLP_LOOP_LEN_RDIV_BY_FACTOR_P ?
2.
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> Here too I think the check should be based on a cached record
> of whether we're using SELECT_VL for this particular loop,
Since only Case 1 is using SELECT_VL, Case 2 && 3 is using MIN_EXPR.
I have no idea how to know whether it is using SELECT_VL.
Could you teach me about this?
>> What happens for loop_lens->length () != 1?
As you can see, When we are using SELECT_VL (case 1), the length will be possible non-vf in non-final iteration.
So here we need to adjust data reference pointer IV by the outcome of SELECT_VL.
The gimple IR is like this:
# vectp_B.8_16 = PHI <vectp_B.8_17(6), &B(5)>
...
_28 = .SELECT_VL (ivtmp_26, POLY_INT_CST [4, 4]);ivtmp_15 = _28 * 4;..vect__1.10_18 = .LEN_LOAD (vectp_B.8_16, 128B, _28, 0);..
vectp_B.8_17 = vectp_B.8_16 + ivtmp_15;
This is the case loop_lens->length () == 1.
When it is loop_lens->length () != 1, it means it is Case 2 or Case 3. We always force MIN_EXPR using VF in non-final iteration.So the data reference IV is added by constant value (poly or non-poly). Maybe the codes here is ugly with using loop_lens->length () == 1?Could you give me the suggestions for this?
I am gonna fix this patch by following your suggestions.Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 00:45
To: juzhe.zhong
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
In addition to Jeff's comments:
juzhe.zhong@rivai.ai writes:
> [...]
> diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
> index cc4a93a8763..99cf0cdbdca 100644
> --- a/gcc/doc/md.texi
> +++ b/gcc/doc/md.texi
> @@ -4974,6 +4974,40 @@ for (i = 1; i < operand3; i++)
> operand0[i] = operand0[i - 1] && (operand1 + i < operand2);
> @end smallexample
>
> +@cindex @code{select_vl@var{m}} instruction pattern
> +@item @code{select_vl@var{m}}
> +Set operand 0 to the number of active elements in vector will be updated value.
> +operand 1 is the total elements need to be updated value.
> +operand 2 is the vectorization factor.
> +The value of operand 0 is target dependent and flexible in each iteration.
> +The operation of this pattern can be:
> +
> +@smallexample
> +Case 1:
> +operand0 = MIN (operand1, operand2);
> +operand2 can be const_poly_int or poly_int related to vector mode size.
> +Some target like RISC-V has a standalone instruction to get MIN (n, MODE SIZE) so
> +that we can reduce a use of general purpose register.
> +
> +In this case, only the last iteration of the loop is partial iteration.
> +@end smallexample
> +
> +@smallexample
> +Case 2:
> +if (operand1 <= operand2)
> + operand0 = operand1;
> +else if (operand1 < 2 * operand2)
> + operand0 = IN_RANGE (ceil (operand1 / 2), operand2);
GCC's IN_RANGE is a predicate, so it would be best to avoid that here.
Why isn't it simply ceil (operand1 / 2), which must be <= operand2?
> +else
> + operand0 = operand2;
> +
> +This case will evenly distribute work over the last 2 iterations of a stripmine loop.
> +@end smallexample
> +
> +The output of this pattern is not only used as IV of loop control counter, but also
> +is used as the IV of address calculation with multiply/shift operation. This allow
> +us dynamic adjust the number of elements is processed in each iteration of the loop.
> +
> @cindex @code{check_raw_ptrs@var{m}} instruction pattern
> @item @samp{check_raw_ptrs@var{m}}
> Check whether, given two pointers @var{a} and @var{b} and a length @var{len},
> [...]
> diff --git a/gcc/tree-ssa-loop-manip.cc b/gcc/tree-ssa-loop-manip.cc
> index 909b705d00d..5abca64379e 100644
> --- a/gcc/tree-ssa-loop-manip.cc
> +++ b/gcc/tree-ssa-loop-manip.cc
> @@ -47,7 +47,9 @@ along with GCC; see the file COPYING3. If not see
> so that we can free them all at once. */
> static bitmap_obstack loop_renamer_obstack;
>
> -/* Creates an induction variable with value BASE + STEP * iteration in LOOP.
> +/* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
> + If CODE is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
> + If CODE is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
> It is expected that neither BASE nor STEP are shared with other expressions
> (unless the sharing rules allow this). Use VAR as a base var_decl for it
> (if NULL, a new temporary will be created). The increment will occur at
> @@ -57,8 +59,8 @@ static bitmap_obstack loop_renamer_obstack;
> VAR_AFTER (unless they are NULL). */
>
> void
> -create_iv (tree base, tree step, tree var, class loop *loop,
> - gimple_stmt_iterator *incr_pos, bool after,
> +create_iv (tree base, tree_code code, tree step, tree var,
> + class loop *loop, gimple_stmt_iterator *incr_pos, bool after,
> tree *var_before, tree *var_after)
> {
> gassign *stmt;
> @@ -66,7 +68,9 @@ create_iv (tree base, tree step, tree var, class loop *loop,
> tree initial, step1;
> gimple_seq stmts;
> tree vb, va;
> - enum tree_code incr_op = PLUS_EXPR;
> + /* The code can only be PLUS_EXPR or MINUS_EXPR. */
> + gcc_assert (code == PLUS_EXPR || code == MINUS_EXPR);
> + tree_code incr_op = code;
As Richard said, we should be able to get rid of incr_op, probably
by calling the parameter incr_op.
I think the later:
incr_op = MINUS_EXPR;
stmts need to be updated to something that flips between PLUS_EXPR
and MINUS_EXPR (with updates to the comments).
It would probably make sense to split the create_iv part out as a
separate prepatch.
> edge pe = loop_preheader_edge (loop);
>
> if (var != NULL_TREE)
> @@ -1365,7 +1369,7 @@ tree_transform_and_unroll_loop (class loop *loop, unsigned factor,
> tree ctr_before, ctr_after;
> gimple_stmt_iterator bsi = gsi_last_nondebug_bb (new_exit->src);
> exit_if = as_a <gcond *> (gsi_stmt (bsi));
> - create_iv (exit_base, exit_step, NULL_TREE, loop,
> + create_iv (exit_base, PLUS_EXPR, exit_step, NULL_TREE, loop,
> &bsi, false, &ctr_before, &ctr_after);
> gimple_cond_set_code (exit_if, exit_cmp);
> gimple_cond_set_lhs (exit_if, ctr_after);
> @@ -1580,8 +1584,8 @@ canonicalize_loop_ivs (class loop *loop, tree *nit, bool bump_in_latch)
> gsi = gsi_last_bb (loop->latch);
> else
> gsi = gsi_last_nondebug_bb (loop->header);
> - create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
> - loop, &gsi, bump_in_latch, &var_before, NULL);
> + create_iv (build_int_cst_type (type, 0), PLUS_EXPR, build_int_cst (type, 1),
> + NULL_TREE, loop, &gsi, bump_in_latch, &var_before, NULL);
>
> rewrite_all_phi_nodes_with_iv (loop, var_before);
>
> [...]
> diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> index 44bd5f2c805..d63ded5d4f0 100644
> --- a/gcc/tree-vect-loop-manip.cc
> +++ b/gcc/tree-vect-loop-manip.cc
> @@ -385,6 +385,48 @@ vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_controls *dest_rgm,
> return false;
> }
>
> +/* Try to use permutes to define the lens in DEST_RGM using the lens
> + in SRC_RGM, given that the former has twice as many lens as the
> + latter. Return true on success, adding any new statements to SEQ. */
Nit: think it's better to talk about "lengths" rather than "lens"
(which ordinarily means something different).
Like Jeff said, this operation is no longer a permute when applied
to lengths (as opposed to masks). And because of that, I don't think
there's anything special about DEST_RGN having twice as many controls
as SRC_RGN. (And the implementation does seem to be more general
than just a factor of two.)
> +
> +static bool
> +vect_maybe_permute_loop_lens (tree iv_type, gimple_seq *seq,
> + rgroup_controls *dest_rgm,
> + rgroup_controls *src_rgm)
> +{
> + tree ctrl_type = dest_rgm->type;
> + poly_uint64 nitems_per_ctrl
> + = TYPE_VECTOR_SUBPARTS (ctrl_type) * dest_rgm->factor;
> + if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter)
> + {
> + for (unsigned int i = 0; i < dest_rgm->controls.length (); ++i)
> + {
> + tree src = src_rgm->controls[i / dest_rgm->controls.length ()];
> + tree dest = dest_rgm->controls[i];
> + gassign *stmt;
> + if (i == 0)
> + {
> + /* MIN (X, VF*I/N) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl);
> + stmt = gimple_build_assign (dest, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + else
> + {
> + /* (X - MIN (X, VF*I/N)) capped to the range [0, VF/N]. */
> + tree factor = build_int_cst (iv_type, nitems_per_ctrl * i);
> + tree temp = make_ssa_name (iv_type);
> + stmt = gimple_build_assign (temp, MIN_EXPR, src, factor);
> + gimple_seq_add_stmt (seq, stmt);
> + stmt = gimple_build_assign (dest, MINUS_EXPR, src, temp);
> + gimple_seq_add_stmt (seq, stmt);
> + }
> + }
> + return true;
> + }
> + return false;
> +}
> +
This return false thing looks odd. I don't think we can decide on
a group-by-group basis whether to use this approach or not. Instead
it has to be a per-loop decision. Either (1) we use SELECT_VL for the
single-control group, and use this function to derive all other groups
from that, or (2) we just MIN for all groups. No mixing and matching
is allowed.
In other words, I think we need to decide at the top level whether
to use SELECT_VL or not. As things stand, this decision should be
"no" if nitems_per_iter != 1 for *any* group. If the decision is
to use SELECT_VL, we should use it for the single-control group
only and then unconditionally use the function above for the other
groups, with no failure path.
The decision about whether to use SELECT_VL or not should be cached
if necessary, so that it's available...
> /* Helper for vect_set_loop_condition_partial_vectors. Generate definitions
> for all the rgroup controls in RGC and return a control that is nonzero
> when the loop needs to iterate. Add any new preheader statements to
> [...]
> @@ -10361,15 +10361,18 @@ vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> }
>
> /* Given a complete set of length LENS, extract length number INDEX for an
> - rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS. */
> + rgroup that operates on NVECTORS vectors, where 0 <= INDEX < NVECTORS.
> + Insert any set-up statements before GSI. */
>
> tree
> -vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
> - unsigned int nvectors, unsigned int index)
> +vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi,
> + vec_loop_lens *lens, unsigned int nvectors, tree vectype,
> + unsigned int index)
> {
> rgroup_controls *rgl = &(*lens)[nvectors - 1];
> bool use_bias_adjusted_len =
> LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0;
> + tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo);
>
> /* Populate the rgroup's len array, if this is the first time we've
> used it. */
> @@ -10400,6 +10403,27 @@ vect_get_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens,
>
> if (use_bias_adjusted_len)
> return rgl->bias_adjusted_ctrl;
> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
> + OPTIMIZE_FOR_SPEED))
> + {
> + tree loop_len = rgl->controls[index];
> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
> + if (maybe_ne (nunits1, nunits2))
> + {
> + /* A loop len for data type X can be reused for data type Y
> + if X has N times more elements than Y and if Y's elements
> + are N times bigger than X's. */
> + gcc_assert (multiple_p (nunits1, nunits2));
> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
> + gimple_seq seq = NULL;
> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
> + build_int_cst (iv_type, factor));
> + if (seq)
> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
> + }
> + return loop_len;
> + }
> else
> return rgl->controls[index];
> }
...here. That is, the key isn't whether SELECT_VL is available,
but instead whether we've decided to use it for this loop (unless
I'm missing something).
> diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> index 3ad6a7d28d7..26813ab56e9 100644
> --- a/gcc/tree-vect-stmts.cc
> +++ b/gcc/tree-vect-stmts.cc
> @@ -3147,6 +3147,61 @@ vect_get_data_ptr_increment (vec_info *vinfo,
> return iv_step;
> }
>
> +/* Prepare the pointer IVs which needs to be updated by a variable amount.
> + Such variable amount is the outcome of .SELECT_VL. In this case, we can
> + allow each iteration process the flexible number of elements as long as
> + the number <= vf elments.
> +
> + Return data reference according to SELECT_VL.
> + If new statements are needed, insert them before GSI. */
> +
> +static tree
> +get_select_vl_data_ref_ptr (vec_info *vinfo, stmt_vec_info stmt_info,
> + tree aggr_type, class loop *at_loop, tree offset,
> + tree *dummy, gimple_stmt_iterator *gsi,
> + bool simd_lane_access_p, vec_loop_lens *loop_lens,
> + dr_vec_info *dr_info,
> + vect_memory_access_type memory_access_type)
> +{
> + loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
> + tree step = vect_dr_behavior (vinfo, dr_info)->step;
> +
> + /* TODO: We don't support gather/scatter or load_lanes/store_lanes for pointer
> + IVs are updated by variable amount but we will support them in the future.
> + */
> + gcc_assert (memory_access_type != VMAT_GATHER_SCATTER
> + && memory_access_type != VMAT_LOAD_STORE_LANES);
> +
> + /* When we support SELECT_VL pattern, we dynamic adjust
> + the memory address by .SELECT_VL result.
> +
> + The result of .SELECT_VL is the number of elements to
> + be processed of each iteration. So the memory address
> + adjustment operation should be:
> +
> + bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + addr = addr + .SELECT_VL (ARG..) * bytesize;
> + */
> + gimple *ptr_incr;
> + tree loop_len
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens, 1, aggr_type, 0);
> + tree len_type = TREE_TYPE (loop_len);
> + poly_uint64 bytesize = GET_MODE_SIZE (element_mode (aggr_type));
> + /* Since the outcome of .SELECT_VL is element size, we should adjust
> + it into bytesize so that it can be used in address pointer variable
> + amount IVs adjustment. */
> + tree tmp = fold_build2 (MULT_EXPR, len_type, loop_len,
> + build_int_cst (len_type, bytesize));
> + if (tree_int_cst_sgn (step) == -1)
> + tmp = fold_build1 (NEGATE_EXPR, len_type, tmp);
> + tree bump = make_temp_ssa_name (len_type, NULL, "ivtmp");
> + gassign *assign = gimple_build_assign (bump, tmp);
> + gsi_insert_before (gsi, assign, GSI_SAME_STMT);
It seems a little wasteful to recompute the temporary results (into gsi)
for every pointer, but I guess it's not worth complicating the code to
do on-the-fly DCE. Later passes can clean it up for us. (Unless Richi
feels differently. Can't remember who's said what now.)
> + return vect_create_data_ref_ptr (vinfo, stmt_info, aggr_type, at_loop, offset,
> + dummy, gsi, &ptr_incr, simd_lane_access_p,
> + bump);
> +}
> +
> /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
>
> static bool
> @@ -8279,7 +8334,7 @@ vectorizable_store (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
> incr = gsi_stmt (incr_gsi);
> @@ -8540,6 +8595,17 @@ vectorizable_store (vec_info *vinfo,
> vect_get_gather_scatter_ops (loop_vinfo, loop, stmt_info,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + simd_lane_access_p ? loop : NULL,
> + offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
Here too I think the check should be based on a cached record
of whether we're using SELECT_VL for this particular loop,
What happens for loop_lens->length () != 1?
Thanks,
Richard
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -8788,8 +8854,9 @@ vectorizable_store (vec_info *vinfo,
> else if (loop_lens)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies, vec_num * j + i);
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> + vec_num * j + i);
> tree ptr = build_int_cst (ref_type, align * BITS_PER_UNIT);
> machine_mode vmode = TYPE_MODE (vectype);
> opt_machine_mode new_ovmode
> @@ -9450,7 +9517,7 @@ vectorizable_load (vec_info *vinfo,
>
> stride_base = cse_and_gimplify_to_preheader (loop_vinfo, stride_base);
> ivstep = cse_and_gimplify_to_preheader (loop_vinfo, ivstep);
> - create_iv (stride_base, ivstep, NULL,
> + create_iv (stride_base, PLUS_EXPR, ivstep, NULL,
> loop, &incr_gsi, insert_after,
> &offvar, NULL);
>
> @@ -9928,6 +9995,16 @@ vectorizable_load (vec_info *vinfo,
> slp_node, &gs_info, &dataref_ptr,
> &vec_offsets);
> }
> + else if (loop_lens && loop_lens->length () == 1
> + && direct_internal_fn_supported_p (
> + IFN_SELECT_VL, LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo),
> + OPTIMIZE_FOR_SPEED)
> + && memory_access_type != VMAT_INVARIANT)
> + dataref_ptr
> + = get_select_vl_data_ref_ptr (vinfo, stmt_info, aggr_type,
> + at_loop, offset, &dummy, gsi,
> + simd_lane_access_p, loop_lens,
> + dr_info, memory_access_type);
> else
> dataref_ptr
> = vect_create_data_ref_ptr (vinfo, first_stmt_info, aggr_type,
> @@ -10144,8 +10221,8 @@ vectorizable_load (vec_info *vinfo,
> else if (loop_lens && memory_access_type != VMAT_INVARIANT)
> {
> tree final_len
> - = vect_get_loop_len (loop_vinfo, loop_lens,
> - vec_num * ncopies,
> + = vect_get_loop_len (loop_vinfo, gsi, loop_lens,
> + vec_num * ncopies, vectype,
> vec_num * j + i);
> tree ptr = build_int_cst (ref_type,
> align * BITS_PER_UNIT);
> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
> index 9cf2fb23fe3..357bc5c7315 100644
> --- a/gcc/tree-vectorizer.h
> +++ b/gcc/tree-vectorizer.h
> @@ -2293,8 +2293,8 @@ extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
> unsigned int, tree, unsigned int);
> extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> tree, unsigned int);
> -extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
> - unsigned int);
> +extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, vec_loop_lens *,
> + unsigned int, tree, unsigned int);
> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
> extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 10:11 ` juzhe.zhong
@ 2023-05-11 11:04 ` Richard Sandiford
2023-05-11 11:21 ` juzhe.zhong
0 siblings, 1 reply; 25+ messages in thread
From: Richard Sandiford @ 2023-05-11 11:04 UTC (permalink / raw)
To: juzhe.zhong; +Cc: gcc-patches, rguenther
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai> writes:
> Hi， Richard. Since create_iv has been approved and soon will be commited after
> we bootstrap && regression.
>
> Now, I plan to send patch for "decrement IV".
>
> After reading your comments, I have several questions:
>
> 1.
>> if (use_bias_adjusted_len)
>> return rgl->bias_adjusted_ctrl;
>> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
>> + OPTIMIZE_FOR_SPEED))
>> + {
>> + tree loop_len = rgl->controls[index];
>> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
>> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
>> + if (maybe_ne (nunits1, nunits2))
>> + {
>> + /* A loop len for data type X can be reused for data type Y
>> + if X has N times more elements than Y and if Y's elements
>> + are N times bigger than X's. */
>> + gcc_assert (multiple_p (nunits1, nunits2));
>> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
>> + gimple_seq seq = NULL;
>> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
>> + build_int_cst (iv_type, factor));
>> + if (seq)
>> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
>> + }
>> + return loop_len;
>> + }
>> else
>> return rgl->controls[index];
>> }
>
>> ...here. That is, the key isn't whether SELECT_VL is available,
>> but instead whether we've decided to use it for this loop (unless
>> I'm missing something).
>
> Let's me clarify it again:
>
> I do this here is for Case 2 SLP:
>
> Generate for len : _61 = _75 / 2;
> I think it is similar with ARM SVE using VIEW_CONVER_EXPR to view_convert the mask.
>
> You said we should not let SELECT_VL is available or not to decide it here.
> Could you teach me how to handle this code here? Should I add a target hook like:
> TARGET_SLP_LOOP_LEN_RDIV_BY_FACTOR_P ?
No. What I mean is: for each vectorised loop, we should make a decision,
in one place only, whether to use SELECT_VL-based control flow or
arithmetic-based control flow for that particular loop. That decision
depends partly on direct_internal_fn_supported_p (a necessary but not
sufficient condition), partly on whether the loop contains SLP nodes, etc.
We should then record that decision in the loop_vec_info so that it is
available to whichever code needs it.
This is similar to LOOP_VINFO_USING_PARTIAL_VECTORS_P etc.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 11:04 ` Richard Sandiford
@ 2023-05-11 11:21 ` juzhe.zhong
2023-05-11 11:29 ` Richard Sandiford
0 siblings, 1 reply; 25+ messages in thread
From: juzhe.zhong @ 2023-05-11 11:21 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 2966 bytes --]
Oh, I see. But I saw there is a variable using_partial_vectors_p
in the loop data structure.
Can I add a variable call using_select_vl_p ?
Since it may increase the size of data structure, I am not sure whether it is appropriate.
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 19:04
To: juzhe.zhong\@rivai.ai
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai> writes:
> Hi， Richard. Since create_iv has been approved and soon will be commited after
> we bootstrap && regression.
>
> Now, I plan to send patch for "decrement IV".
>
> After reading your comments, I have several questions:
>
> 1.
>> if (use_bias_adjusted_len)
>> return rgl->bias_adjusted_ctrl;
>> + else if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type,
>> + OPTIMIZE_FOR_SPEED))
>> + {
>> + tree loop_len = rgl->controls[index];
>> + poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type);
>> + poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype);
>> + if (maybe_ne (nunits1, nunits2))
>> + {
>> + /* A loop len for data type X can be reused for data type Y
>> + if X has N times more elements than Y and if Y's elements
>> + are N times bigger than X's. */
>> + gcc_assert (multiple_p (nunits1, nunits2));
>> + unsigned int factor = exact_div (nunits1, nunits2).to_constant ();
>> + gimple_seq seq = NULL;
>> + loop_len = gimple_build (&seq, RDIV_EXPR, iv_type, loop_len,
>> + build_int_cst (iv_type, factor));
>> + if (seq)
>> + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT);
>> + }
>> + return loop_len;
>> + }
>> else
>> return rgl->controls[index];
>> }
>
>> ...here. That is, the key isn't whether SELECT_VL is available,
>> but instead whether we've decided to use it for this loop (unless
>> I'm missing something).
>
> Let's me clarify it again:
>
> I do this here is for Case 2 SLP:
>
> Generate for len : _61 = _75 / 2;
> I think it is similar with ARM SVE using VIEW_CONVER_EXPR to view_convert the mask.
>
> You said we should not let SELECT_VL is available or not to decide it here.
> Could you teach me how to handle this code here? Should I add a target hook like:
> TARGET_SLP_LOOP_LEN_RDIV_BY_FACTOR_P ?
No. What I mean is: for each vectorised loop, we should make a decision,
in one place only, whether to use SELECT_VL-based control flow or
arithmetic-based control flow for that particular loop. That decision
depends partly on direct_internal_fn_supported_p (a necessary but not
sufficient condition), partly on whether the loop contains SLP nodes, etc.
We should then record that decision in the loop_vec_info so that it is
available to whichever code needs it.
This is similar to LOOP_VINFO_USING_PARTIAL_VECTORS_P etc.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 11:21 ` juzhe.zhong
@ 2023-05-11 11:29 ` Richard Sandiford
2023-05-11 12:08 ` juzhe.zhong
0 siblings, 1 reply; 25+ messages in thread
From: Richard Sandiford @ 2023-05-11 11:29 UTC (permalink / raw)
To: juzhe.zhong; +Cc: gcc-patches, rguenther
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai> writes:
> Oh, I see. But I saw there is a variable using_partial_vectors_p
> in the loop data structure.
>
> Can I add a variable call using_select_vl_p ?
Yeah. Please also add a wrapper macro like
LOOP_VINFO_USING_PARTIAL_VECTORS_P. (I'm not really a fan of the
wrappers, but it's better to be consistent.)
> Since it may increase the size of data structure, I am not sure whether it is appropriate.
The structure is only temporary, and very few of them exist at
a given time. Besides, there's already a layout hole on LP64 hosts
around those booleans (between slp_unrolling_factor and scalar_loop).
So the new boolean shouldn't grow the size of the structure.
We can convert the booleans to bitfields if size ever becomes a problem.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 11:29 ` Richard Sandiford
@ 2023-05-11 12:08 ` juzhe.zhong
2023-05-11 12:42 ` Richard Sandiford
0 siblings, 1 reply; 25+ messages in thread
From: juzhe.zhong @ 2023-05-11 12:08 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 1455 bytes --]
Thanks. I have read rgroup descriptions again.
Still I am not fully understand it clearly, bear with me :)
I don't known how to differentiate Case 2 and Case 3.
Case 2 is multiple rgroup for SLP.
Case 3 is multiple rgroup for non-SLP (VEC_PACK_TRUNC)
Is it correct:
case 2: rgc->max_nscalarper_iter != 1
Case 3 : rgc->max_nscalarper_iter == 1 but rgc->factor != 1?
Thanks.
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 19:29
To: juzhe.zhong\@rivai.ai
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai> writes:
> Oh, I see. But I saw there is a variable using_partial_vectors_p
> in the loop data structure.
>
> Can I add a variable call using_select_vl_p ?
Yeah. Please also add a wrapper macro like
LOOP_VINFO_USING_PARTIAL_VECTORS_P. (I'm not really a fan of the
wrappers, but it's better to be consistent.)
> Since it may increase the size of data structure, I am not sure whether it is appropriate.
The structure is only temporary, and very few of them exist at
a given time. Besides, there's already a layout hole on LP64 hosts
around those booleans (between slp_unrolling_factor and scalar_loop).
So the new boolean shouldn't grow the size of the structure.
We can convert the booleans to bitfields if size ever becomes a problem.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 12:08 ` juzhe.zhong
@ 2023-05-11 12:42 ` Richard Sandiford
2023-05-11 23:12 ` 钟居哲
0 siblings, 1 reply; 25+ messages in thread
From: Richard Sandiford @ 2023-05-11 12:42 UTC (permalink / raw)
To: juzhe.zhong; +Cc: gcc-patches, rguenther
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai> writes:
> Thanks. I have read rgroup descriptions again.
> Still I am not fully understand it clearly, bear with me :)
>
> I don't known how to differentiate Case 2 and Case 3.
>
> Case 2 is multiple rgroup for SLP.
> Case 3 is multiple rgroup for non-SLP (VEC_PACK_TRUNC)
>
> Is it correct:
> case 2: rgc->max_nscalarper_iter != 1
Yes.
> Case 3 : rgc->max_nscalarper_iter == 1 but rgc->factor != 1?
For case 3 it's:
rgc->max_nscalars_per_iter == 1 && rgc != &LOOP_VINFO_LENS (loop_vinfo)[0]
rgc->factor is controlled by the target and just says what units
IFN_LOAD_LEN works in. E.g. if we're loading 16-byte elements,
but the underlying instruction measures bytes, the factor would be 2.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
* Re: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
2023-05-11 12:42 ` Richard Sandiford
@ 2023-05-11 23:12 ` 钟居哲
0 siblings, 0 replies; 25+ messages in thread
From: 钟居哲 @ 2023-05-11 23:12 UTC (permalink / raw)
To: richard.sandiford; +Cc: gcc-patches, rguenther
[-- Attachment #1: Type: text/plain, Size: 1155 bytes --]
Hi, Richards.
Could you take a look at this patch:
https://gcc.gnu.org/pipermail/gcc-patches/2023-May/618241.html
Thanks
juzhe.zhong@rivai.ai
From: Richard Sandiford
Date: 2023-05-11 20:42
To: juzhe.zhong\@rivai.ai
CC: gcc-patches; rguenther
Subject: Re: [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai> writes:
> Thanks. I have read rgroup descriptions again.
> Still I am not fully understand it clearly, bear with me :)
>
> I don't known how to differentiate Case 2 and Case 3.
>
> Case 2 is multiple rgroup for SLP.
> Case 3 is multiple rgroup for non-SLP (VEC_PACK_TRUNC)
>
> Is it correct:
> case 2: rgc->max_nscalarper_iter != 1
Yes.
> Case 3 : rgc->max_nscalarper_iter == 1 but rgc->factor != 1?
For case 3 it's:
rgc->max_nscalars_per_iter == 1 && rgc != &LOOP_VINFO_LENS (loop_vinfo)[0]
rgc->factor is controlled by the target and just says what units
IFN_LOAD_LEN works in. E.g. if we're loading 16-byte elements,
but the underlying instruction measures bytes, the factor would be 2.
Thanks,
Richard
^ permalink raw reply [flat|nested] 25+ messages in thread
end of thread, other threads:[~2023-05-11 23:12 UTC | newest]
Thread overview: 25+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-05-04 13:25 [PATCH V4] VECT: Add decrement IV iteration loop control by variable amount support juzhe.zhong
2023-05-05 23:41 ` 钟居哲
2023-05-09 12:59 ` Richard Sandiford
2023-05-09 13:27 ` juzhe.zhong
2023-05-09 14:34 ` 钟居哲
2023-05-07 15:19 ` Jeff Law
2023-05-07 21:54 ` 钟居哲
2023-05-09 10:52 ` juzhe.zhong
2023-05-08 5:35 ` Kewen.Lin
2023-05-08 6:27 ` juzhe.zhong
2023-05-08 7:55 ` Kewen.Lin
2023-05-08 8:25 ` juzhe.zhong
2023-05-10 16:45 ` Richard Sandiford
2023-05-10 21:00 ` 钟居哲
2023-05-10 21:28 ` Richard Sandiford
2023-05-10 22:51 ` 钟居哲
2023-05-11 4:50 ` Richard Sandiford
2023-05-11 5:14 ` juzhe.zhong
2023-05-11 10:11 ` juzhe.zhong
2023-05-11 11:04 ` Richard Sandiford
2023-05-11 11:21 ` juzhe.zhong
2023-05-11 11:29 ` Richard Sandiford
2023-05-11 12:08 ` juzhe.zhong
2023-05-11 12:42 ` Richard Sandiford
2023-05-11 23:12 ` 钟居哲
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).